AI vs AI: Cybersecurity in the Age of Generative Threats

AI对决AI：生成式威胁时代下的网络安全 随着Anthropic的Mythos等先进AI模型的应用，攻击者能够迅速将多个低风险漏洞串联为高危害的攻击链，令受监管限制而难以即时修补的医疗IoT等设备面临严峻挑战。由于AI驱动的攻击可在数分钟或数秒内发生，传统数周的响应周期已不再适用，防守方必须采用自动化防御控制，并在高风险操作中保留“人在回路”与确定性的硬护栏。与此同时，专家提醒需警惕对AI的盲目信任和第三方API单点故障等盲点，推动企业安全治理从单纯的书面行政政策转向由技术护栏直接落地执行。 攻击者利用先进LLM快速将低影响漏洞串联为高危害攻击链，而受制于FDA再认证等监管要求的传统医疗IoT设备无法即时打补丁，迫使防守方高度依赖补偿性控制。 AI的执行速度将传统数周的响应SLA压缩至数分钟，促使防守方采用自动化隔离手段，并通过高风险需人工授权、低风险自动化执行的分级矩阵来平衡效率与安全。 过度依赖算法会导致人类监督能力弱化，且单一依赖外部LLM API存在停机等系统性风险，这要求组织将行政合规流程现代化并转化为自动化的技术护栏。

As generative AI accelerates the discovery, chaining, and execution of cyberattacks, security leaders argue that defense must transition from slow administrative policies to automated technical controls while maintaining hard guardrails against over-reliance. AI-Accelerated Attack Velocity and Vulnerability Chaining Adversaries are leveraging advanced models, such as Anthropic's Mythos, to compress attack timelines from weeks down to minutes or seconds. Rather than relying solely on single critical flaws, bad actors use generative tools to rapidly identify multiple low-impact, low-risk vulnerabilities and chain them into high-impact attack vectors. Furthermore, AI automates the transition from scanning targets to actively deploying payloads across non-traditional entry points, including HVAC systems, medical office workstations, and operational technology (OT) networks. Compensating Controls and the Realities of Medical IoT Medical internet-of-things (IoT) and OT systems present acute operational challenges because known outdated software often cannot be patched immediately due to FDA re-certification mandates. Consequently, security teams must deploy compensating controls around these static endpoints rather than applying direct updates. Alaa Abdulridha notes that while models like Mythos assist defenders by discovering flaws, models trained on outdated or corrupted data frequently hallucinate and produce false positives. In this environment, human analysts remain necessary to validate automated findings and remediate real flaws. Balancing Autonomous Defense With Economic and Operational Risk Automated response tools can counter machine-speed threats by quarantining compromised hosts or isolating non-critical network segments before human review. Torrell Funderburk cites a benchmark by Kevin Mandia's firm where an autonomous defensive AI successfully neutralized all 38 attack paths mapped across 17 million events by an offensive AI. However, Funderburk cautions against an 'infinite-token' defense, noting that attempting to defend every asset against low-cost, high-volume AI assaults exhausts budgets and introduces operational risks. A mistaken automated change to a core firewall or production pipeline can cause more disruption than the original threat. Systemic Blind Spots and Technical Guardrails Relying on autonomous security introduces psychological and architectural vulnerabilities. Krista Arndt warns that unquestioned trust in AI erodes human intuition and oversight over time. On an architectural level, organizations risk operational failure by relying heavily on cloud-hosted LLM APIs from vendors like OpenAI or Anthropic without maintaining manual fallbacks during outages. Panelists advocate a risk-tiered operational model: low-risk tasks execute automatically, while high-risk actions demand human authorization backed by immutable, deterministic rules that block catastrophic actions, such as blanket open firewall configurations. Shifting Governance From Paperwork to Enforcement Traditional Governance, Risk, and Compliance (GRC) workflows that take weeks to draft manual policies fail against rapid automated attacks. While generative tools can draft policies in seconds, panel members argue that written rules cannot stop users from making mistakes. Effective governance must instead embed technical enablers directly into networks to enforce compliance programmatically. At St. Luke's University Health Network, AI implementations must pass cross-functional clinical, ethical, and business alignment reviews before technical teams assess their cybersecurity risks. While automated defenses have proven effective in benchmarked simulations, organizations still face the unresolved dilemma of managing single-point dependencies on third-party AI APIs and preventing severe self-inflicted operational disruptions caused by automated remediation.

随着生成式AI将漏洞挖掘与攻击执行的周期压缩至以秒计算，防御体系面临传统响应窗口失效与复合漏洞串联的双重挑战。与会专家指出，企业必须建立结合确定性技术护栏的风险分级自动化防御，同时警惕模型幻觉、云端API单点依赖以及对自主决策的过度信任。 攻击提速与漏洞串联重塑防御响应 以Anthropic的Mythos为代表的先进AI模型，正在根本性地改变网络攻防的速度与战术。传统安全服务等级协议（SLA）通常允许数天至三周的响应窗口，但在攻击者利用AI实现秒级利用的情况下，这种周期已经彻底失效。Stephen Franklin指出，围绕Mythos的宣传虽包含市场炒作，但其核心威胁在于执行速率；AI工具不仅能像Nmap一样发现开放端口，更能自主完成从识别到向边缘目标推送利用程序的全流程。攻击范围也从传统核心服务器蔓延至暖通空调（HVAC）控制器、医疗诊所闲置工作站等非传统端点，正如Colonial Pipeline事件所揭示的次要路径风险。 此外，先进模型善于挖掘代码中的次要瑕疵，并将多个低风险漏洞迅速串联为高危攻击链。在医疗物联网（IoT）与运营技术（OT）场景中，该问题尤为棘手：由于美国FDA等机构严格的重新认证要求，许多已知存在漏洞的老旧设备无法即时安装补丁，安全主管只能依赖补偿性控制措施。对此，NetWatch.AI等防御方案通过自动化手段将可疑主机隔离或移出关键网段，随后将高优先级告警移交人工安全运营中心（SOC）。Alaa Abdulridha强调，防御方应将此类模型视为自动化重复性工作的放大器，由人类专家主导验证与修复，而非完全由模型替代分析人员。 自主防御的边界与确定性风险护栏 在探讨是否应授予AI自主修改核心防火墙规则或推送生产补丁的权限时，专家们评估了自主攻防对抗的成效与代价。Torrell Funderburk援引了Kevin Mandia机构的一项基准测试：在历时三天、涵盖1700万次事件和238项发现的实测中，全自主进攻AI绘制了38条有效攻击路径，但最终全部被自主防御AI成功拦截。这一结果验证了算法化实时阻断的技术可行性。 然而，Funderburk警告企业不能将安全防御演变为成本高昂的“无限Token博弈”。试图调用高额算力不加区分地拦截所有低成本、高批量的AI探测，会迅速耗尽企业预算并带来沉重的运营负担；更为严重的是，一旦自主AI在核心网络上产生误判（例如误封关键业务流量），造成的损失可能远超漏洞本身。因此，企业需要构建结构化风险矩阵：针对高影响或高风险操作强制保留人工介入（Human-in-the-loop），仅对低风险环节开放完全自主代理工作流。同时，网络配置必须设置硬性确定性护栏，如硬编码严禁生成全网开放（0.0.0.0/0）等高危规则，且医疗等特定领域的安全开支必须始终与患者护理和临床运营目标保持对齐。 算法盲区、单点依赖与过度信任风险 AI安全工具的技术局限与组织心理盲区带来了深层系统性隐患。Krista Arndt提出，“盲目信任”是长期的重大心理脆弱点，人类对自主决策系统的依赖可能在2040年左右迎来临界点；当安全人员习惯于默认算法正确时，关键的人类直觉与审计质疑能力将被严重削弱。在疲劳和注意力分散状态下，用户无法始终做出明智的安全抉择，这使得底层机制的设计必须依赖技术控制而非人类自律。 在架构与技术层面，过度依赖Anthropic或OpenAI等第三方LLM供应商的云端API，构成了致命的单点故障风险。一旦外部API服务中断或底层架构遭遇零日漏洞，那些已经裁撤传统手动运维能力的企业将丧失应急备用手段。此外，第三方厂商普遍缺乏对训练数据集透明度的公开，采购方只能依据输出表现进行黑盒评估。Abdulridha补充道，虽然LLM在能直接读取源代码的白盒代码审查中表现优异，但在缺乏上下文信息的黑盒渗透测试中表现极为吃力，加之训练数据过时引发的幻觉与虚假告警，企业依然需要人工专家进行常态化校验。 现代治理：从行政纸面合规转向技术落地 传统的治理、风险与合规（GRC）模式依赖庞大团队耗费数周撰写行政安全规范，这种工作节奏已无法应对AI时代的安全需求。虽然Microsoft Copilot等生成式工具可在30秒内起草一套制度文本，但单纯的纸面禁令并不能阻止违规行为发生。Arndt指出，现代企业治理必须从行政文本转向“技术化治理”，即将合规策略直接编译并固化在自动化的技术护栏中，通过系统机制强制实施约束，例如通过自动化链接拦截来防范用户点击恶意钓鱼网站。 在落地实践中，St. Luke's University Health Network构建了多层级的企业治理管线。在技术团队介入安全测试之前，所有拟引入的AI系统必须先通过跨职能委员会的审查，该委员会首先从组织商业模式、患者健康影响以及机构文化等维度评估工具的临床价值与伦理合规性。只有在通过了临床和伦理维度的严格筛查后，项目才会进入深度的网络安全与技术风险评估环节，以此确保技术部署在提速的同时不偏离医疗机构的核心安全底线。 生成式AI在重塑网络攻防速度的同时，也迫使安全架构从滞后的行政合规与人工响应转向基于确定性护栏的技术化治理。然而，在自动化对抗的演进过程中，如何平衡算力经济成本、防范模型幻觉误判，并在第三方云端模型发生中断时保留有效的手动降级控制，仍是企业在深度部署自主防御前必须解决的关键矛盾。

## Background Ai4 频道于 2026 年 8 月 21 日在 YouTube 平台上发布了题为《AI vs AI: Cybersecurity in the Age of Generative Threats》的圆桌研讨会视频。该内容汇集了来自医疗、金融科技和前沿安全初创公司的多位 CISO 与技术专家，共同探讨在攻击者与防守方均引入大模型的背景下，企业如何应对 AI 生成式钓鱼、Deepfake 等新型威胁并构建自动化防御与治理体系。 ## TL;DR 本次研讨深入剖析了生成式 AI 浪潮下攻防对抗的演进逻辑，指出攻防的核心差异在于攻击速度与并发规模的指数级提升，而非单点扫描能力的质变。来自医疗、金融与安全软件领域的专家分享了自主 SOC 建设、全网微隔离、多通道通信验真及用户行为基线建模等一线防御实践。嘉宾们强调，单纯依靠无止境消耗 Token 无法打赢防御战，企业必须通过硬编码的确定性红线与“人在回路”（Human-in-the-Loop）机制控制 AI 的自主修复权限以防止系统性失效。最后，讨论呼吁摆脱繁重低效的传统行政治理，转向技术驱动的敏捷防御，并坚守身份分层、邮件防伪协议等扎实的基础安全防线。 ## Atmosphere & style 整场研讨会呈现出务实、犀利且极具一线实战气息的专业基调，嘉宾们没有停留在泛化的概念宣传上，而是直接切入大模型落地中的幻觉、误报与供应链痛点。主持人与嘉宾之间的互动默契流畅，既有对医疗与金融等高严苛行业治理难点的深刻剖析，也穿插着对行业盲目跟风“AI 炒作”的幽默调侃。现场氛围理性而笃定，从技术对抗逐渐升华至对安全第一性原理的思考，带给听众一种既紧跟前沿又回归工程常识的踏实感。 ## Key takeaways - [01:49](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=109s) 传统安全标准已无法完全覆盖 AI 引入的非传统控制项，企业需结合 NIST AI RMF、OWASP LLM 威胁清单及厂商情报自建定制化治理模型。 - [05:35](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=335s) 底层数据质量与时效性是 AI 安全智能体的命脉，失真或滞后的数据会被大模型成倍放大并直接引发严重幻觉与误报漏洞。 - [08:37](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=517s) 对抗 Deepfake 与生成式社工钓鱼不仅需要多通道通信验真，更依赖基于用户日常系统操作轨迹建立的行为基线（Baselines）进行异常偏离拦截。 - [10:36](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=636s) 面对 AI 生成的高拟真钓鱼攻击，企业防御应回归基础管理，通过身份分层隔离（如服务台权限限制）及关键财务流程的四眼双人复核来化解风险。 - [18:16](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=1096s) 大模型攻防的核心差异在于攻击执行速度（从发现到 Payload 部署被压缩至数秒），传统的 SLA 响应机制失效，防守方必须赋予防御 AI 实时的自动化阻断与自愈权限。 - [21:29](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=1289s) AI 攻击方擅长跨边界扫描 IT、OT、边缘终端及云服务的每一个薄弱环节，防御方必须构建全域统一的覆盖体系以避免不对称对抗劣势。 - [22:47](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=1367s) 在近乎无限并发的攻击 AI 面前，防御方单纯依靠燃烧 Token 全面硬抗会导致经济成本崩溃，必须通过资产范围缩减（Descoping）聚焦保护高价值核心业务。 - [26:31](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=1591s) 为防范自治系统的单点灾难性故障，安全架构必须将核心不可逾越红线写入确定性代码（如禁止 0.0.0.0/0 放行规则），并对高风险变更保留人机协同（Human-in-the-Loop）。 - [27:31](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=1651s) 未来 AI 安全最大的盲区在于过度盲目信任与缺乏失效模式（Failure Mode）兜底，若缺乏人工备用流程，一旦上游大模型 API 遭遇宕机或 0-day 漏洞将造成系统性停摆。 - [33:32](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=2012s) 基于网络心理学（Cyber Psychology）揭示的人性不可靠性，技术手段治理（硬性技术护栏与自动化工具）必须优先于传统的行政规章与免责文档宣导。 - [39:46](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=2386s) 无论技术如何演进，以身份为核心的基础安全基石（如 SPF/DMARC 协议、网络微隔离、非人类身份治理与 mTLS）依然是决定性的根本防线。 - [44:24](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=2664s) 在大模型逐渐大宗商品化（Commoditized）的趋势下，安全对抗的核心瓶颈在于组织内部的“行动授权”（Authority to Act），而非单纯的算力或数据。 - [45:30](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=2730s) 警惕 AI 的过度概念炒作，在许多安全场景下确定性 Python 脚本或 RPA 的效率与稳定性远胜大模型，应以业务成效为导向构建务实防护框架。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=0s) 圆桌嘉宾介绍与开场背景 本场圆桌聚焦于“AI 对抗 AI：生成式威胁时代下的网络安全”，汇集了来自医疗卫生网络、大型金融科技公司以及初创安全软件领域的从业者与企业领袖。主持人 Parthasarathi Chakraborty 率先展开自我介绍，他目前担任 Broadridge Financial Solutions 的副总裁兼全球安全工程主管。Broadridge 是一家大型金融科技企业，而 Parthasarathi 整个职业生涯均深耕于高度受监管的金融（如摩根大通 JPMorgan、美国银行 Bank of America）与医疗保险领域（如 Humana）。此外，他还兼任西北大学（Northwestern University）网络安全硕士项目的兼职讲师。 随后，现场嘉宾由右至左依次进行了简短的背景介绍： - **Alaa Abdulridha**：SerpApi 的工程总监兼网络安全研究员，全面负责 SerpApi 内部的合规体系与网络安全保障。 - **Stephen Franklin**：监控解决方案提供商 NetWatch.AI 的创始人兼 CEO；该公司专注于监控类产品，并在圆桌前一周刚刚发布了全新的 Agentic（智能体）产品。 - **Torrell Funderburk**：Overspace 的创始人兼 CEO，此前曾两度出任跨国企业全球 CISO；Overspace 专注于网络弹性（Cyber Resilience），主攻业务流程测绘（Business Process Mapping）与巨灾建模（Catastrophe Modeling）。 - **Krista Arndt**：圣卢克大学健康网络（St. Luke's University Health Network）的副 CISO。该医疗组织规模庞大，位于宾夕法尼亚州的里海谷（Lehigh Valley），旗下管辖 16 个主要院区及 400 多个分支机构。 --- ### [01:49](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=109s) 医疗网络的 AI 治理困局与自主 SOC 实践 进入正式议题，主持人 Parthasarathi 向 Krista 提问：在圣卢克这样的庞大医疗机构中，排名前一至两位的 AI 网络安全重点举措是什么？在立项时又遵循何种方法论或决策框架？Krista 指出，当前最核心的两大项目是**自主安全运营中心（Autonomous SOC）**与**威胁管理自动化**，同时团队正在落地类似 Gartner 在五年前提出的“网络安全网格架构”（Cybersecurity Mesh Architecture），实现全网遥测数据（Telemetry）的相互打通、交叉关联与数据富化。 ```mermaid flowchart TD accTitle: 医疗网络安全与AI治理框架架构 subgraph Regulatory["合规基准与传统框架"] HIPAA["HIPAA 医疗合规法规"] NIST_CSF["NIST 传统网安框架"] end subgraph Emerging["新兴标准与威胁情报"] AIRMF["NIST AI RMF"] OWASP["OWASP LLM 威胁清单"] ISO["ISO 规范"] VendorIntel["厂商发现与同行情报"] end subgraph CustomFramework["定制化内部治理模型"] Custom["自研非传统控制项标准"] SecGov["自建网络安全治理模型"] EntGov["企业级 AI 采纳治理"] end HIPAA --> Custom NIST_CSF --> Custom AIRMF --> Custom OWASP --> Custom ISO --> Custom VendorIntel --> Custom Custom --> SecGov SecGov --> EntGov ``` 在框架与方法论方面，Krista 坦言，作为严格受 HIPAA 法规管制的医疗实体，机构一直将 NIST 视作具备法律抗辩力（defensible）的基石。然而面对当下 AI 引入的大量非传统控制项，传统的 NIST 框架已捉襟见肘。尽管业界涌现了 NIST AI RMF、OWASP 针对大语言模型的威胁清单，以及 ISO 发布的 LLM 安全构建指南，“但没有任何单一标准能够完整覆盖我们正在摸索的非传统控制领域”。为此，Krista 最近牵头结合了厂商安全发现、行业信息共享以及 OWASP 威胁清单，撰写了一套机构专用的非传统控制项标准，并配套建立了网络安全治理模型与企业治理策略，以此来规范全院“究竟该不该用 AI、该如何用 AI”。 - 主持人追问该自研框架是否对外公开，Krista 爽快表示非常愿意在去标识化（de-identify）处理后公开共享。 - 该框架在内部推进时得到了高层领导与知名全国性权威机构的认可与背书，其 CISO 更是凭借卓越的战略眼光争取到了充足支持。 - Krista 谦虚地补充：“我并不是说我们做到了十全十美，但我很乐意共享出来；或许能帮到某些同行，或者大家能指出我们哪里还能改进，毕竟我也不可能通晓所有事情。” --- ### [04:06](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=246s) 从跨国制造到弹性初创：聚焦业务价值与自研自用 随后，主持人向 Torrell 提问：从前跨国 CISO 到如今创立 Overspace，他在主导 AI 驱动的安全项目时有何重心与决策逻辑？Torrell 回顾了自己此前在跨国制造巨头希悦尔（Sealed Air）担任全球 CISO 的经历。希悦尔在全球遍布 150 多个工业站点，几乎覆盖所有国家。这段经历带给他最深刻的教训是：企业往往面临“在所有地方同时做所有事情（everything everywhere all at once）”的诱惑，但这种撒胡椒面式的做法极度低效。 在跨国制造业，核心决策依据是“如何支持营收运营（Revenue Operations）”。因此，当时所有的 AI 安全举措、攻击面加固和应急响应，全部紧密锚定在保障营收流程的范畴内。在创立 Overspace 后，Torrell 将这种务实的哲学延续到了初创企业中： - 决策标准依然是清晰界定“对企业究竟什么才重要”。 - 作为一家网络弹性公司，首要原则就是“吃自己的狗粮”（dog food our own product），拿自身产品和基础设施做试验田。 - 团队在引入或对抗 AI 时，始终围绕底层风险展开推演：“可能发生在我们身上的最坏情况是什么？AI 在其中能提供什么支撑，反过来 AI 又会构成哪些毁灭性威胁？” --- ### [05:35](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=335s) 降低使用门槛的智能体与防御模型的数据保真挑战 NetWatch.AI 的 CEO Stephen 随后阐述了其公司今年的两大战略核心。第一项重头戏是刚刚上线的 Agentic（智能体）产品，核心目标是彻底解决普通用户的 Prompt 门槛问题。许多非技术用户因担心提示词写错导致系统给出错误答案，而对 AI 心存芥蒂甚至拒绝使用。NetWatch.AI 研发了针对高频场景的交互机制，让智能体在感知环境异动时主动发起探测并向用户提问。第二大方向则是布局企业内部的“AI 治理（AI Governance）”——不仅要监控自身智能体的健康度，还要对企业环境中运行的其他外部 AI 模型进行跨域监督与管控，该功能目前正处于规模化扩展阶段。 ```mermaid flowchart LR accTitle: SerpApi 数据质量保障与防幻觉链条 subgraph Ingestion["外部数据捕获"] WebData["全网非结构化数据"] --> Engine["结构化搜索引擎 API"] end subgraph QualityControl["数据校验与质量防线"] Engine --> Clean["数据清洗与源头核验"] Clean --> Verify{"时效性与完整性校验"} end subgraph Consumer["安全智能体消费"] Verify -- 合格数据 --> SecModels["漏洞挖掘 / 合规审计 (ISO等)"] Verify -. 污染/陈旧数据.-> Risk["模型严重幻觉 / 误报爆发"] end ``` 紧接着，来自 SerpApi 的 Alaa 从核心数据服务商的角度分享了截然不同的技术挑战。由于 SerpApi 为行业内大量 AI 软件及下游智能体提供连接互联网和结构化搜索引擎的通道，Alaa 团队所面临的核心痛点是**数据的准确性、时效性与数据完整性（Data Integrity）**： - 底层数据的微小瑕疵都会被 AI 模型成倍放大，直接引发模型严重幻觉（Hallucination）。 - 在网络安全下游场景中，若数据源头质量低下或更新滞后，安全分析类 Agent 在漏洞挖掘（Vulnerability Discovery）中就会产生海量误报（False Positives）。 - 对于用于自动化合规审计（如 ISO 认证合规）的专有 Agent，陈旧或失真数据更是会直接摧毁整个自动化合规流程的有效性。 --- ### [07:38](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=458s) 金融科技视角：安全融入业务与架构审查自动化 在各位嘉宾分享之后，主持人 Parthasarathi 结合自身在 Broadridge 的实践补充了金融科技大厂的权衡取舍。Broadridge 业务覆盖 27 个国家，在历史发展中累计收购了多达 67 家子公司，系统异构性极强。在这样的环境下推进 AI，安全团队确立的首要标准是：**绝不能让安全成为 AI 业务创新的阻碍（Blocker）**。 在业务团队追求极速交付的压力下，安全团队必须做到“在不造成伤害（do no harm）的前提下将安全无缝内嵌”。Parthasarathi 强调，过去任何新项目上线生产环境，都必须经历耗时费力的安全架构审查（Security Architecture Review），这种传统的人工流程漫长而痛苦。因此，Broadridge 当前的重头戏就是借助 AI 实现架构安全审查的流程自动化，用效率提升换取业务增速，在安全与业务赋能之间达成动态平衡。 --- ### [08:37](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=517s) 对抗生成式社工钓鱼与 Deepfake：从多通道验真到行为基线 探讨深入到攻击面变化：生成式 AI 的平民化导致高度拟真的钓鱼信息与 Deepfake 欺诈呈现井喷之势。Parthasarathi 询问 Stephen，作为产品研发方，NetWatch.AI 在向客户交付方案前，采取了哪些切实可行的对抗手段？ ```mermaid flowchart TD accTitle: 多通道通信验真与行为基线防御 subgraph Channels["外联与通信验真机制"] SMS["短信 SMS"] --> Carrier["运营商认证与官方 Logo 贴标 (Verizon/AT&T)"] Email["告警邮件"] --> ColorMask["专有邮件色彩与格式掩码 (Color-coding)"] Voice["电话语音"] --> PauseDetect["AI 合成语音特征识别 (停顿分析/机械音检测)"] end subgraph Intelligence["行为基线分析 (AI 治 AI)"] Telemetry["7天用户操作与应用行为数据"] --> Baseline["建立个人日常操作节奏 (Rhythm)"] Baseline --> Anomaly{"偏离常规基线?"} Anomaly -- 是 --> Trigger["拦截告警 / 识破假冒会话"] Anomaly -- 否 --> Pass["合法放行"] end ``` Stephen 坦言，系统发出的告警如果仅仅通过普通电子邮件分发，极难向用户证明其真实性，而短信（SMS）也早已成为钓鱼重灾区。对此，他们采取了一系列多维度的刚性验证手段： - **运营商官方认证**：联合 Verizon 和 AT&T，推进发件方权威核验；当告警短信送达客户手机时，会直接强制显示官方品牌 Logo。 - **色彩与模式掩码（Color-coded Masking）**：在告警邮件中引入特有的色彩方案；若伪造者不知晓系统当下的动态色彩编码机制，用户即可迅速甄别伪造邮件。 - **电话语音合成鉴别**：针对越来越逼真的 AI 语音拟声电话，Stephen 幽默地提示可以“等待那个停顿”——常与 AI 打交道的人会发现合成音往往存在微妙的机械停滞或泛化模板感；目前团队正推进经过认证的固定电话（Verified Landline）连接机制。 当主持人追问这是否本质上是“用 AI 对抗 AI（Handling AI with AI）”时，Stephen 给出了极为核心的底层逻辑：**行为基线（Baselines）**。NetWatch.AI 花费大量精力在客户环境中采集底层数据，通常仅需约 7 天时间，系统即可根据应用使用轨迹和个人操作规律勾勒出用户的“日常节律”（Rhythm）——例如用户几点活跃、何时喝咖啡休息（并非通过物理摄像头，而是纯粹从系统交互行为推断）。一旦后续交互行为严重偏离这一固有节奏，即便攻击者伪造了天衣无缝的音色或话术，防御模型也能瞬间捕捉到反常并进行阻断。 ### [10:36](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=636s) 回归基础防御：身份分层与小团队的抗钓鱼之道 当安全运营从小规模扩展到企业级时，虽然可以通过 SIEM 或传统监控方案标记风险行为，但引入 AI 后能够真正捕获并串联起全量数据点。在谈及如何以初创企业有限的团队规模抵御 Deepfake（深度伪造）和 AI 赋能的钓鱼攻击时，Torrell Funderburk 坦言，自己的应对策略听起来甚至有些反直觉，因为这是一种“很多人并不想听到的基础方法（basics）”，但它却具备极高的可扩展性。 Torrell 指出，上市公司面临的最主要威胁之一就是针对服务台（Help Desk）的社会工程学与钓鱼攻击。对此，最有效的防御手段并非堆砌复杂工具，而是做好基础的身份架构管理： - **身份分层与隔离（Identity Tiering & Segmentation）**：严格限制服务台人员权限，例如一线支持人员绝对无权直接修改特权账户（Privileged Identity）的凭据。 - **关键业务的双人复核机制（Four-Eyes Principle）**：对所有涉及资金划转和核心财务流程的操作强制推行双人审批规则。 针对日益难以肉眼识别的 AI 生成式钓鱼邮件，Alaa Abdulridha 补充说明，由于黑客与普通人都在使用大语言模型，单纯依靠内容语义判断真伪已不再可靠。SerpApi 正在探索利用其大规模实时搜索数据，构建一种类似于“背景调查”的 AI Agent。该 Agent 能对邮件或来电的来源进行全网数据验证，并生成一个可信度评分（Rating Score），从而快速判定其为合法通信还是钓鱼威胁。 ```mermaid flowchart TD accTitle: AI时代的基础防御与来源验证机制 subgraph S1["身份与流程基础控制"] A["服务台/Help Desk"] -->|"权限隔离限制"| B["禁止直接修改特权凭据"] C["财务/敏感流程"] -->|"强制执行"| D["四眼原则 / 双人审批"] end subgraph S2["AI 来源可信度核验"] E["可疑邮件 / 来电"] --> F["全网实时数据检索 Agent"] F --> G["生成可信度评分 (Rating Score)"] G --> H{"判定通信合法性"} end ``` --- ### [13:22](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=802s) 医疗关键基础设施保护：微隔离、网络心理学与身份验证 面对利用 LinkedIn 等社交网络进行精细化情报侦察（Recon）的精准社会工程学攻击，医疗行业面临着关乎患者生命安全的极端挑战。Krista Arndt 分享了圣卢克大学健康网络（St. Luke's University Health Network）在安全架构上的突破性进展——架构团队在短短 46 天内从头到尾（soup to nuts）落地了全网颗粒度的微隔离（Granular Micro-segmentation）。在医疗环境中，微隔离不仅要保障手术辅助机器人等高精度自治设备的网络接入与隔离，还要对安全性普遍较弱且受制于 FDA 冗长认证周期的生物医疗设备（Biomedical Devices）构筑坚固防线。 在人员防线方面，Krista 强调了网络心理学（Cyber Psychology）的应用价值。医疗工作者在高压环境下承担着多线程任务，虽然主观上极具安全合规意愿，但认知负荷极易导致疏漏。因此，机构正全面升级身份验证体系，推行包括医师生物识别、服务台双向身份核验，以及针对非人类身份（Non-Human Identities）的重点治理，以配合 AI 驱动的业务迭代速度。 Parthasarathi Chakraborty 结合自身在大型机构（如 Humana 等物联网/医疗设备密集型企业）的工程管理经验，总结了纵深防御的工程化落地要点： - **严格推行邮件防伪协议**：强制将 DMARC 策略设置为严格拒绝（`p=reject`），结合 SPF 严防域名仿冒（Domain Spoofing）与接管。 - **抗钓鱼多因素认证**：部署符合 FIDO 标准、具备防劫持/防中间人攻击能力的抗仿冒 MFA。 - **带外验证机制（Out-of-Band Validation）**：在 AI 能够轻松伪造屏幕信息与声音的时代，针对可疑请求与凭据变更触发带外人工核验流程。 --- ### [18:16](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=1096s) 应对前沿大模型威胁：Mythos、AI 提速与自动化攻防对抗 针对 Anthropic、OpenAI 等前沿大模型（如 Mythos 等自动化漏洞挖掘与利用模型）进入攻击者视野的趋势，讨论转向了如何应对“以 AI 速度发起的威胁”。当攻击者利用大模型将原本耗时数周的漏洞链组合压缩到极短时间内，甚至把多个低危缺陷级联成高危利用链时，传统的安全响应时效已面临失效。 Alaa Abdulridha 认为，虽然类似 Mythos 的模型在漏洞检测方面极具威力，但如果依赖过时的网络数据，模型同样会产生幻觉（Hallucinations）和误报（False Positives）。因此，防御侧应将此类工具定位为赋能工具，用于接替安全专家的重复性事务，缩短发现、验证及排除误报的时间周期，而非完全取代专业安全人员。 Stephen Franklin 则指出，部分关于 Mythos 类前沿工具的宣传存在一定的市场营销成分。在 NetWatch.AI 的实际测试中，虽然大模型在单点脆弱性识别上表现出色，但其本质能力并未完全超出传统扫描工具（如 Nmap）的范畴。攻防对抗的核心差异在于**攻击执行的速度（Speed of Attack）**——即从发现漏洞到自动化部署 Payload 的时间窗口被极度压缩。面对几分钟甚至几秒钟级的攻击周期，“3 周甚至 2 小时的 SLA 响应已成历史”，防御方必须赋予防御型 AI 与攻击方等同的自主执行权限，实现实时的自动化风险阻断与系统自愈。 ### [21:29](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=1289s) 全面防御与跨网络边界的 AI 威胁覆盖 在谈及威胁处置流程时，安全团队的核心逻辑是尽早识别异常行为，并在必要时先执行隔离或阻断操作——例如将受感染主机置于离线状态、切断其网络连接或直接剥离应用访问权限，随后在数分钟内自动向分析师、服务台（Help Desk）或安全运营中心（SOC）派发告警工单进行人工介入。许多企业在保护核心业务资产（高优先级目标）方面表现出色，但往往会忽视边缘资产，例如医疗机构内未及时打补丁的工作站、电力公司推迟更新的终端设备，或者是运营技术（OT）与暖通空调系统（HVAC）。 当年 Colonial Pipeline 管道公司遭受勒索攻击正是从这类边缘系统切入的。现代高级威胁模型（如 Mythos 风格的攻击模型）最擅长从整个信息孤岛中全面搜寻薄弱环节。如果企业的网络层防御固若金汤，但某个 SaaS 应用或云服务存在漏洞，攻击者就能顺藤摸瓜完成渗透。由于攻击方利用 AI 进行全盘无死角的扫描与攻击，防守方也必须采用全方位的统一视角，建立覆盖 IT、OT 以及云环境的整体防御体系，否则必将在不对称对抗中处于下风。 ```mermaid flowchart TD accTitle: 全面防御与响应流程图 subgraph AssetCoverage["全域资产覆盖与监测"] A["传统 IT / 高优先级目标"] B["边缘终端 / 医疗工作站"] C["OT / 暖通 HVAC / 物联网"] D["云端服务与 SaaS 应用"] end subgraph ThreatEngine["AI 威胁分析与响应"] E["统一威胁分析引擎 (AI Engine)"] F{"风险评估与处置"} G["自动隔离 / 断网 / 撤销权限"] H["人工研判 (SOC / 帮助台 / 分析师)"] end A --> E B --> E C --> E D --> E E --> F F -->|"高危/紧急"| G F -->|"数分钟内派发"| H ``` ### [22:47](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=1367s) 自主修复的信任边界与攻防经济学：攻防 AI 模拟实战 当讨论延伸至是否应将关键基础设施的修复与补丁控制权完全移交给 AI 时，大型金融机构面临着巨大的信任与合规挑战。以摩根大通（JPMorgan）或 Broadridge 这种拥有数百台大型防火墙的企业为例，一旦 AI 误配防火墙规则，造成的客户业务中断将是灾难性的。在探讨此类自主修复模式时，专家们指出了攻击与防御背后的经济学问题：防御方真正担忧的往往不是单点模型的能力上限，而是 AI 带来的攻击规模与并发量的指数级激增。 前不久 Kevin Mandia（Mandiant 创始人）新成立的公司公布了一项受控实验：一场完全自主的进攻性 AI 与完全自主的防御性 AI 之间的红蓝对抗测试。该测试在三天内连续生成了 1700 万次事件，识别出 238 项攻击发现并持续映射攻击生命周期。最终，在 1700 万次攻击尝试中绘制出的 38 条完整攻击路径全部被自主防御 AI 成功拦截。 - **对抗规模与消耗**：三天内 1700 万次攻击事件、238 项发现、映射出 38 条攻击路径，且防御方达成全量拦截。 - **Token 消耗与防御成本**：在低成本攻击 AI 发起的近乎无限次并发攻击面前，如果防守方单纯依靠无止境地燃烧 Token 进行全面对抗，最终会在经济成本上率先崩溃。 - **降维策略与聚焦**：企业不能盲目追求“在所有地方同时做所有事”（everything everywhere all at once），必须采取本地化 AI 辅助内部配置管理，或对资产进行激进的范围缩减（descoping），全力保护核心高价值业务。 ### [26:31](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=1591s) 风险分级与确定性代码：防御控制权的分层委托 对于依赖自治系统的潜在风险，如果自动化链条中某一环出现故障，可能导致整个安全体系瞬间瘫痪。因此，安全控制权必须遵循基于风险的分级管理模型：对于高影响、高风险的变更场景，必须保留“人类在回路中”（Human-in-the-Loop）的审批机制；而对于低影响、低风险的操作，则可以放手实行自动化。尤其在医疗等行业，安全团队的核心预算与精力必须优先保障患者护理和医疗创新，安全体系的构建不能脱离业务的终极使命。 为了在给予 AI 执行权限的同时规避灾难性故障，安全架构中必须引入硬编码的确定性规则（Deterministic Code）。例如，防火墙变更中坚决禁止配置开放给所有地址（`0.0.0.0/0`）的放行规则，终端检测与响应（EDR）工具也不得擅自关闭关键防护组件。通过将企业前十大潜在灾难性误操作写死在底层的不可逾越红线中，既能让 AI 承担日常琐碎运维，又从根本上防止了“交出王国钥匙”带来的系统性毁灭。 ### [27:31](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=1651s) 未来数年 AI 在安全领域的核心盲区：过度信任与故障模式 在探讨未来几年 AI 安全应用的最大盲区时，与会嘉宾不约而同地聚焦于对技术的“盲目信任”与系统失效后的应对断层。AI 固然极大地提升了威胁管理效率，弥补了分析师在时间与精力上的客观局限，但随着对智能工具的依赖加深，信任奇点（Singularity with Trust）的隐患逐渐浮现。从控制论与网络心理学（Cyberpsychology）的角度看，最大的隐患在于人类过早、完全地将关键决策全盘托付给 AI。 ```mermaid flowchart TD accTitle: AI 安全盲区与依赖风险架构 subgraph BlindSpots["AI 安全核心盲区"] T["盲目信任 (Blind Trust)<br/>过早放弃人工研判"] F["单点失效模式 (Failure Modes)<br/>缺乏模拟兜底流程"] D["上游依赖风险 (Upstream Zero-day)<br/>大模型 API 中断与漏洞"] end subgraph BusinessImpact["业务连续性影响"] C1["业务流程完全停摆"] C2["灾难恢复 (DR) 失效"] C3["供应链穿透风险"] end T --> C1 F --> C2 D --> C3 ``` 企业往往只设计了 AI 处于正常工作状态下的“顺畅路径”（Green Path），却未曾规划 AI 崩溃时的“失效模式”（Failure Mode）。正如近期某些大型云端 API 突发宕机时，许多依赖自动化调用的企业客户瞬间陷入瘫痪，因为原先熟悉业务流程的人工团队早已缩编或流失，企业根本没有准备好备用的手动模拟操作机制。一旦 Anthropic、OpenAI 或底层云服务提供商遭遇 0-day 漏洞或服务中断，将直接演变为企业不可承受的灾难性单点故障。 ### [30:07](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=1807s) 渗透测试局限与训练数据质量：AI 模型的供应链与黑盒挑战 从攻防实战和产品工程的角度来看，当前的 AI 模型在完全取代专业安全人员方面仍存在显著的技术短板。在渗透测试领域，AI 在黑盒测试（Blackbox Testing）中的表现依然较弱，容易陷入低效的重复性试错；相反，在白盒测试（Whitebox Testing）和源代码审计（Code Review）中，由于可以直接获取底层代码上下文，AI 的表现则相当出色。此外，模型的输出上限完全受制于输入数据的质量，劣质的数据训练必然导致劣质的推理决策。 关于大模型提供商的第三方及第四方供应链风险，当前大多数头部 AI 厂商普遍对其预训练数据的具体来源与清洗过程缺乏透明度。虽然安全团队在引入供应商时倾向于要求其提供全面的尽职调查与来源清单，但在当下的实际商业落地中，获取绝对透明的数据血统难度极大。现阶段的工程重心往往更多地落在评估模型本身的输出质量、鲁棒性以及在具体安全场景下的防护表现上。 ### [32:23](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=1943s) 数据质量的竞争与传统治理模式的失效 研讨会继续探讨了 AI 模型的底层核心要素——数据。高质量的真实数据直接决定了模型的优劣，而各家 AI 厂商当前竞争的真正焦点正是高品质数据的获取来源。随后，主持人将话题转向组织内部的 AI 治理（Governance）、信任建立与人机协同（Human-in-the-loop）。 主持人指出，传统的治理模式已经无法适应当前的 AI 演进速度： - 过去大型企业通常需要数十上百人的 GRC（治理、风险与合规）团队耗费数周时间起草一份合规政策。 - 如今，借助 Copilot 等 AI 工具，30 秒即可生成 20 条现成策略。 - 在医疗等强监管行业中，甚至需要为每一项微小事务制定政策声明，这种传统冗长而僵化的管理体系已经完全不可持续。 ```mermaid flowchart LR accTitle: 治理模式对比：传统GRC vs AI时代技术治理 subgraph Traditional["传统合规体系（不可持续）"] A["百人GRC团队"] --> B["耗时数周撰写文档"] B --> C["纯行政规章/免责条款"] C --> D["依赖人工遵守 (易失效)"] end subgraph Modern["AI驱动现代治理（技术优先）"] E["AI自动化辅助"] --> F["技术治理与护栏"] F --> G["多层综合评估链条"] G --> H["全生命周期持续监控"] end ``` --- ### [33:32](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=2012s) 技术性治理优先于行政规章：人性的不确定性与严格评估链 作为医疗机构的首席信息安全官（CISO），Krista Arndt 分享了她在跨行业转型中的治理洞察。她曾先后供职于美国国防部（DoD）、银行业、加密货币领域，最终来到医疗行业。她笑称自己原本以为加密领域会更难，但医疗体系的治理复杂性更为前沿。银行业的经验让她深刻意识到：**必须在所有合理的地方全面推行增强手段，并坚持以风险为导向。** Krista 强调了行政治理（Administrative Governance）与技术治理（Technical Governance）的本质差异： - **行政治理的局限性**：许多人谈到治理就只想到行政管理政策。尽管行政策略具备法律责任明确和树立先例的价值，但现实中“不要点击可疑链接”的政策即使每天宣导，人们依然天天误点。 - **技术治理的必要性**：面对网络心理学（Cyber Psychology）所揭示的人性弱点——员工可能会感到疲惫、心情不好，或是存在有意/无意的内部威胁——人类永远不可能完全做到绝对可靠与一致。因此，必须将“技术手段治理”置于首位，通过自动化防护工具和技术护栏来硬性消减高危风险。 - **全链条业务评估**：在引入 AI 方面，企业设立了严密的治理体系：首先筛选该项目是否契合商业模式、企业文化与患者利益，权衡其正向收益是否压倒潜在风险；随后才移交至安全治理与风险审查环节，辅以临床层面的技术护栏，确保 AI 技术全面契合组织整体诉求。 --- ### [35:39](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=2139s) 软件供应链与第三方治理的信任机制 针对初创企业面对大型企业严苛合规审计的挑战，Torrell Funderburk 分享了从产品构建与第三方风险管理（TPRM）角度的应对策略。主持人提出，许多企业依赖 AI 工具来自动生成 SOC 2 等合规证明，但这造成了“先有鸡还是先有蛋”的悖论：企业无法确知背后的第三方/第四方 LLM 供应商自身究竟采用了何种治理标准。 Torrell 指出： - **回归第一性原理**：SOC 2 等合规认证并不直接等同于拥有完美的治理，它传达的核心信号是“该团队正在努力以正确的方式做事”。 - **透明度与韧性共享**：客户真正需要的是稳定可靠的服务。因此，团队不仅要阐明构建服务与保护产品的方法，还专门建立了一套持续信任机制，向第三方实时公开企业的韧性状态与潜在失效风险，使评估透明化。 --- ### [37:42](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=2262s) 敏捷团队的 AI 人才培养与真实数据训练 对于初创公司而言，在资源有限的条件下跟进前沿技术并培养团队技能是一大挑战。Stephen Franklin 和 Alaa Abdulridha 分享了各自的应对实践： - **高校顶尖人才发掘与实测验证**：Stephen Franklin 提到，他们在设有办事处的城市（如北卡罗来纳州夏洛特以及肯尼亚内罗毕）直接深入大学的 AI 课堂，寻找排名前两位的顶尖学生提供实习岗位，以此捕捉最新学术趋势并控制用人成本。同时，针对市面上频繁发布的新模型（例如 DeepSeek、Menthas 等），实验室会设立统一测试基准进行严格压测，滤除公关营销水分。 - **依托海量真实结构化数据**：Alaa Abdulridha 表示，作为全球领先的搜索数据提供商，他们重点关注那些使用最高质量、最精准数据的 AI 模型。目前没有任何 AI 可以完全替代专业人员，但经过大规模、真实世界且高度结构化的数据训练的模型表现最为可靠。 --- ### [39:46](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=2386s) 核心总结：AI 时代的攻防对抗与网络安全基石 在进入开放问答前，主持人对核心讨论进行了全面梳理： 1. **数据即货币（Data is the Currency）**：在借助 AI 实现自动化之前，必须优先完成数据清洗并确保底层数据的准确性。 2. **AI 对抗 AI 但保留制衡**：应对高频自动化的 AI 攻击必须以 AI 为刃，但绝不能对 AI 防御彻底放权，必须设置硬编码的规则边界与关键环节的人工监督（Human in the loop）。 3. **技术赋能的敏捷治理**：传统依赖人力堆叠的 GRC 模式无法应对现有规模，治理体系本身必须借助 AI 提速，同时要求第三方与第四方技术供应链具备高度的透明度。 4. **底层安全基石始终不变**：无论是应对 Deepfake 还是新型邮件钓鱼，底层的安全防护手段依然是决定性防线： - 基础身份与通信校验协议：SPF、DKIM、DMARC 等； - 纵深防御架构：网络微隔离（Micro-segmentation）； - 以身份为核心的边界安全：严格管控非人类身份（Non-human Identity）、缩短生存时间（TTL）以及全面落地双向 TLS（mTLS）。 5. **保持理性，重在落地执行**：面对层出不穷的新模型无需过度恐慌。这就像医学检查中做全身扫描（Full-body Scan）可能会查出数以万计的微小异常，关键在于遵循专业判断，依靠坚实的基础安全实践将过去停留在文档或架构层面的策略真正落地执行。 --- ### [42:22](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=2542s) 现场互动：医疗与政府环境下的政策落地困境 在最后的问答环节，一位在医疗健康行业深耕 6 年的前端/DevSecOps 资深工程师向 Krista 提问。他结合自己在医疗及政府部门的工作经历表示，政府和强监管机构的政策制定与更新通常极其缓慢，但在 AI 快速迭代的当下，传统漫长冗重的安全政策制定方式正面临巨大现实挑战。 ### [42:53](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=2573s) 积极参与行业组织以反哺并影响监管政策 现场观众针对合规与监管提出了痛点问题：许多监管机构制定的政策往往极其被动，呈现出“让你怎么做你就怎么做，却不解释为什么”的僵化状态。观众询问现场嘉宾，在推动或影响这些内外部治理机构（如医疗领域的 Medicare、Medicaid，以及 FDA 等）时，是否有过成功的经验。 ```mermaid flowchart TD accTitle: 影响行业监管与政策制定的闭环路径 subgraph Industry["一线实践经验沉淀"] A["一线 CISO / 安全实践者"] --> B["发现不合理政策提案<br/>(如对中小机构强制网络分段)"] end subgraph Advocacy["行业协会与协调组织"] B --> C["加入协调委员会<br/>(如 HSCC / AMA 等)"] C --> D["建立与监管机构直接沟通通道<br/>(HHS / 华盛顿特区立法者)"] end subgraph Impact["重塑政策走向"] D --> E["反馈真实行业承受力与生存困境"] E --> F["优化 HIPAA 等行业合规法案"] end ``` 针对这一提问，Krista Arndt 深有感触地回应道：“医疗行业确实是一场狂野的旅程（a wild ride）。”从金融科技转行到医疗的她指出，无论身处哪个垂直行业，最有效的策略就是深度参与行业协调组织。在医疗领域，她加入了医疗行业协调委员会（Health Sector Coordinating Council, HSCC），该组织能直接与美国卫生与公众服务部（HHS）对话，甚至协助引导 HIPAA 法案的演进与修订。 她举例说明了这种参与的必要性：当监管部门提议“对所有机构强制推行网络分段（segmentation for everybody）”时，她立刻指出，那些勉强雇得起一个既懂 IT 又懂安全人员的中小型诊所根本无法承受高昂的改造成本。这种脱离实际的政策会直接伤害医疗生态，让原本就资源匮乏的小型机构雪上加霜。Krista 强调，从业者必须将自己的实践经验转化为对政策制定者的游说力量，积极去华盛顿特区与监管者当面沟通，扮演“前线 CISO”的智囊角色。 - **垂直行业参与范例**：银行业可依托 AMA 等行业协会，医疗行业可借助 HSCC 直接与 HHS 沟通。 - **政策反馈核心**：防止脱离实际的合规标准拖垮资源有限的中小机构。 - **职业建议**：将志愿参与政策制定与监管沟通作为职业生涯的重要投入。 --- ### [44:24](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=2664s) 大模型商品化浪潮下的真正瓶颈：算力、遥测还是行动授权？ 现场观众抛出了一个颇具深度的问题：Hugging Face CEO 曾提到中国开源大模型与美国前沿实验室（Frontier Labs）的能力差距正在迅速缩小。如果模型质量逐渐演变为“大宗商品化（commoditized）”，那么未来的 AI 军备竞赛是否会转向“算力（compute）”、“遥测数据（telemetry）”与“行动授权（authority to act）”之争？在这三者中，哪一个才是真正的核心瓶颈？ ```mermaid flowchart LR accTitle: AI 军备竞赛中的三大竞争要素与瓶颈分析 subgraph Drivers["模型能力大宗商品化后的竞争维度"] C["算力<br/>(Compute)"] T["遥测数据<br/>(Telemetry)"] A["行动授权<br/>(Authority to Act)"] end A --> Bottleneck["【核心瓶颈】<br/>组织缺乏 AI 政策与决策权<br/>监管严重滞后于技术发展"] ``` 嘉宾们一致认为“行动授权（authority to act）”是当前最严峻的瓶颈。现场嘉宾指出，他们在日常接触的大量客户中，许多企业甚至连基础的 AI 使用政策都没有出台，更不用说授权系统去决定使用哪个模型或采取自动化防御行动了。 嘉宾借用加密货币（Crypto）的监管历史作为对比：“如果当年在加密领域我们一直干等着政府监管机构跟上节奏，那这个行业根本不可能活下来。”在 AI 领域同样如此，面对飞速演进的威胁，防御者不能一味被动等待政策与权限的审批，很多时候必须遵循“先斩后奏、事后请求原谅（ask for forgiveness）”的原则，勇于在安全防御中建立决策与行动权。 --- ### [45:30](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=2730s) 去对齐/无限制模型与务实防御：拒绝概念炒作，回归确定性成效 针对观众关于“去对齐/无限制模型（obliterated models/uncensored models）在防御场景中的应用价值”这一提问，Alaa Abdulridha 率先发表了看法。他认为，当前的大多数 AI 模型在防御和代码审计（code reviews）方面表现出色，但在进攻性用途上表现并不理想。更重要的是，他重申了 AI 在网络安全中的定位：“AI 并不是用来取代专业安全人员的，AI 的真正使命是替代那些重复性的枯燥工作（repetitive work）。” ```mermaid flowchart TD accTitle: 务实的安全架构决策模型 Goal["业务安全目标 / 确定性成效"] --> Check{"方案选型评估"} Check -->|"确定性规则/标准流程"| Simple["基础脚本 / Python / RPA<br/>(低成本、高效、可靠)"] Check -->|"海量数据研判/代码辅助"| AI["AI 大模型<br/>(作为外围补充工具)"] Simple --> Success["稳定可靠的安全运营体系"] AI --> Success ``` 随后，现场嘉宾针对 AI 实施架构与行业浮躁风气展开了精辟的总结： 1. **构建以成效为导向的确定性框架**：真正具有百万美元价值（million-dollar approach）的方案，是建立一个高度聚焦于业务成效的严密控制框架（harness），然后将 AI 作为辅助手段补充进去，而不是盲目追求“无限生成”。 2. **警惕 AI 过度炒作**：Parthasarathi Chakraborty 幽默地回顾了十年前的行业狂热——当年每位 CIO 的明星项目都是“上云（go to cloud）”，但如果追问“上云究竟要达成什么”，得到的回答往往是“走着瞧，我们边走边看”。如今 AI 领域正在重演这种盲目跟风。很多场景下，简单的 RPA、确定性的 Python 脚本或基础自动化程序，其产出效率和稳定性远高于耗费巨大算力资源的 AI 模型。 嘉宾们最后呼吁，安全团队必须保持头脑清醒，远离概念炒作，把精力放在技术能产生真正实效的领域。 ## Notable quotes > [01:49](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=109s) 没有任何单一标准能够完整覆盖我们正在摸索的非传统控制领域。 — Krista Arndt > [04:06](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=246s) 在所有地方同时做所有事情（everything everywhere all at once）。 — Torrell Funderburk > [07:38](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=458s) 绝不能让安全成为 AI 业务创新的阻碍（Blocker）。 — Parthasarathi Chakraborty > [10:36](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=636s) 很多人并不想听到的基础方法（basics）。 — Torrell Funderburk > [18:16](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=1096s) 3 周甚至 2 小时的 SLA 响应已成历史。 — Stephen Franklin > [33:32](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=2012s) 必须在所有合理的地方全面推行增强手段，并坚持以风险为导向。 — Krista Arndt > [39:46](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=2386s) 数据即货币（Data is the Currency）。 — Parthasarathi Chakraborty > [42:53](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=2573s) 医疗行业确实是一场狂野的旅程（a wild ride）。 — Krista Arndt > [45:30](https://www.youtube.com/watch?v=Er-Zuj3Wfhs&t=2730s) AI 并不是用来取代专业安全人员的，AI 的真正使命是替代那些重复性的枯燥工作（repetitive work）。 — Alaa Abdulridha ## Mentioned Krista Arndt, Parthasarathi Chakraborty, Torrell Funderburk, Stephen Franklin, Alaa Abdulridha, Kevin Mandia, St. Luke's University Health Network, Broadridge Financial Solutions, Overspace, NetWatch.AI, SerpApi, Sealed Air, JPMorgan, Bank of America, Humana, Colonial Pipeline, Anthropic, OpenAI, Hugging Face, Verizon, AT&T, NIST CSF, NIST AI RMF, OWASP, ISO, HIPAA, FDA, HSCC, HHS, AMA, Mythos, DeepSeek, Nmap ## Video description Krista Arndt - CISO, St. Luke's University Health Network Parthasarathi Chakraborty - VP - Global Head of Security Engineering, Broadridge Financials Torrell Funderburk - Founder & CEO, Overspace Stephen Franklin - CEO / Co-Founder, NetWatch.AI Alaa Abdulridha - Engineering Director and Cybersecurity Researcher, SerpApi, LLC As defenders adopt AI to detect and respond to threats faster than ever, attackers are using the same tools to launch increasingly complex, automated, and adaptive attacks. This panel explores the arms race between AI-driven defense and AI-powered offense. Cybersecurity leaders will share how they’re integrating AI into detection and response workflows, what emerging threats they’re tracking, and how they’re preparing their teams to deal with adversarial LLMs, deepfakes, and AI-generated phishing. Learn how to defend your organization when the enemy is just as smart. [Source](https://www.youtube.com/watch?v=Er-Zuj3Wfhs)

[1:49] Krista Arndt: 没有任何单一标准能够完整覆盖我们正在摸索的非传统控制领域。 [4:06] Torrell Funderburk: 在所有地方同时做所有事情（everything everywhere all at once）。 [7:38] Parthasarathi Chakraborty: 绝不能让安全成为 AI 业务创新的阻碍（Blocker）。 [10:36] Torrell Funderburk: 很多人并不想听到的基础方法（basics）。 [18:16] Stephen Franklin: 3 周甚至 2 小时的 SLA 响应已成历史。 [33:32] Krista Arndt: 必须在所有合理的地方全面推行增强手段，并坚持以风险为导向。 [39:46] Parthasarathi Chakraborty: 数据即货币（Data is the Currency）。 [42:53] Krista Arndt: 医疗行业确实是一场狂野的旅程（a wild ride）。 [45:30] Alaa Abdulridha: AI 并不是用来取代专业安全人员的，AI 的真正使命是替代那些重复性的枯燥工作（repetitive work）。

[0:00] Here [0:04] in this panel we have awesome panelists like  we have builders we are producing the products [0:11] we have CISO from healthcare network and we have  all the tech people and uh just a quick introduce [0:17] myself and then I will hand it over to the fellow  panelists here my name is part chakraarti I am [0:22] the global head of security engineering at uh  broadidge financials we are a fintech company and [0:28] all throughout my carrier I have been in either  financials JP or bank of America or uh healthcare [0:35] Humana and uh outside of my broadidge engagement I  also teach as an adjunct lecturer at Northwestern [0:42] University uh for their cyber security masters  program so not in a particular order maybe I will [0:47] go from right and then left I will let Allah  my co-panelist to introduce himself and then [0:53] we will go over thank you hi everyone my name is  Allah I'm engineering director at SER API. Also, [1:00] I'm security researcher at SER API, responsible  for all the compliance and cyber security. Um, and [1:06] um, yeah, uh, feel free to visit SER API to learn  more about what we do and what we provide. Thank [1:12] you, Stephen. Hi, I'm Steven Franklin, CEO,  founder of Netwatch AI. Uh, we are software [1:18] provider from monitoring solutions and, uh, we  just launched our Gentic products uh, last week. [1:23] Torell Thunderberg. I'm a twotime global CISO and  now founder and CEO of Overspace. We are a cyber [1:30] resilience company focused on business process  mapping and catastrophe modeling. Christa. Hi, [1:36] my name is Christa. Uh I am an associate CISO at  St. Luke's University Health Network. We're like [1:41] 16 campuses, 400 plus sites. We're in the Lehigh  Valley. So massive operation. Thank you very much. [1:49] So we can go I mean we do have some planned uh  topics to talk about. So maybe if you can start [1:54] from left to right you're first Christa. So if  you have to list our top one or two cyber security [2:00] AI initiatives in your organization what will be  those? And the follow-up question will be is there [2:06] any methodology or framework that you use when you  choose okay this is the project I'm going to take [2:11] this year. Um so so top initiatives in cyber  security are um autonomous sock and uh threat [2:19] management. So we we just implemented something  for autonomous threat management as well. Um and [2:24] then we are doing the whole like crossorrelation  of telemetry that you know Gartner came out with [2:28] like 5 years ago like the cyber mesh framework  everything talks to everything and then enriches [2:32] telemetry. Um so so we use core framework on all  of our projects. to as being HIPPA regulated. Um, [2:40] we use NIST because it's a defensible framework  and obviously with the non-traditional controls [2:44] out now with AI, NIST doesn't quite cut it.  They have the AI RMF and then you have like [2:48] OASP and yay, ISO says here's how you'd build  an LLM securely, right? But nothing covers the [2:53] gamut of like really non-traditional controls that  we're learning about. And so I actually recently [2:58] just wrote our own framework and standard based  um to cover the non-traditionals um beyond just [3:04] our core critical acumen based on all of this  cool information sharing some stuff that came [3:09] from vendors and their findings and like the  OAS threats and all that other fun stuff. So [3:13] we are now using a conglomerate of that as well  as um our cyber security governance model that we [3:18] built inhouse and then also our enterprise  governance to say like what what should we [3:22] even use and and do with AI. Yeah, just quick  question. the custom framework that you use in [3:27] your placement. Is it publicly available? Can it  be shared and people can use it? I would be happy [3:31] to kind of deidentify it and share our our what  we have learned that we've had a lot of support [3:36] from some really reputable national organizations.  Um my CISO is freaking amazing from a strategic [3:41] perspective and so he got support for that and  so they even blessed him. They were like, "Oh, [3:45] you guys like like spot on." And so we we have our  technical um I'm not saying everything is perfect, [3:51] but we have our technical acumen covered.  Yeah. So I I would be happy to um deidentify [3:55] and share at least maybe it'll help someone or  you'll be able to tell me how to do it better [3:59] because I don't know everything. Excellent. So  Terrell former CEO sorry and now CEO founder. So [4:07] uh you are building product and helping. So from  your perspective in your current role that you are [4:13] building. So what will be your top cyber security  AIdriven initiatives and what was your decisioning [4:21] criteria? Why did you choose those? Yeah. So,  um what we're doing now kind is informed by [4:27] uh my background as an actual global CISO. Um uh  before I founded Overpace, I was a global CESO [4:32] at Seed Air, which is a global manufacturing  organization, 150 sites globally across every [4:38] country you can think of. Um and the one thing  I learned from that experience is uh there's an [4:45] opportunity to do everything everywhere all at  once and it's not effective. So like how do we [4:50] choose what's effective? how do we choose what's  most val valuable? Um and for seair it was more [4:56] um how do we support uh revenue operations? So  all of our AI efforts were kind of um anchored on [5:02] revenue operations and then from a security and  response and hardening standpoint it was still [5:07] focused on um scope to revenue operations. Um so  when I when uh I founded overpace uh that same [5:13] approach like what matters to us and one we're  a cyber resilience company. Um so the best thing [5:19] that we could do in terms of applying AI is uh  dog food our own product. So we measure ourselves [5:25] based on our product um our infrastructure like  what's the worst thing that could happen to us [5:30] um and how can AI support that and how does AI  threaten that. Got it. Thank you. So Stephen [5:36] if I can ask the same question to you. Um yeah  so the the first two things that we worked on [5:42] um this year for AI would be launching agentic  agents making it usable for every user who's not [5:49] um familiar with prompting right like the part of  the issue I think with AI right now some people [5:53] who who haven't used it or doesn't like it or  whatever worried they ask it the wrong thing they [5:57] get the wrong information so we postulated ways  to do frequently asked questions the way you can [6:02] have the the AI agentic say okay this is something  going on in the environment kind of probe the user [6:06] um that was number one it does work really  Well, uh, and number two, uh, we're working [6:10] on AI governance as a as a piece of that. So,  not just watching ourselves, but watching other [6:15] EIs in the environment. And so, that that's  something that we're scaling right now. Um, [6:19] and so that I think that that's going to be number  two for this year. Got it. Thank you. And for you, [6:24] what will be your top one or two uh initiatives  that you guys are working on? Um, yeah. So as you [6:30] know that we provide the most of the AI companies  and AI agents with access to the structured search [6:37] engines and data uh from the internet and the  biggest challenge for us was the accuracy and [6:43] the quality of these data. Um because when once  the data is bad the AI model or agent will start [6:51] hallucinating and um it will start making issues.  Um and it was very important for us to keep these [6:59] data very updated and to revise the quality of  these data uh and the integrity and make sure [7:05] the data has very high quality from the source  to be provided to these AI models especially the [7:11] uh the security AI models and agents. Um any any  small uh issue with this data or bad quality uh [7:18] will result for uh a lot of false positives when  it comes to vulnerability discovery for example or [7:25] uh outdated data uh that is being used for agents  that are used for uh compliance like um so ISO and [7:34] so so on. Got it. So yeah I mean like we all have  like our top one or two initiatives. I mean I can [7:40] also add like from broad ridge like what we are  looking at right now. So our decisioning criteria [7:46] is that number one make sure like security is  not becoming a blocker for the AIdriven all of [7:53] the initiative because business wants to go fast  and how can we bring security in line with that [7:58] that's the number one priority and again at the  same time do no harm and also uh how can we ensure [8:04] like security is actually embedded. So some of the  initiatives where it was like more time consuming [8:09] thing like we are an enterprise with uh operating  in 27 countries and grown up with 67 companies [8:17] acquitting 67 companies. So any new initiative  which goes in production has to go through a [8:22] rigorous security architecture review and those  were like painful manual process time-conuming [8:27] thing. So how can we automate like those are the  things where we can gain efficiency and also uh [8:32] keep security in the mind like when we are doing  AI enablement. So those are some of the things we [8:37] also consider. So if I can u move on a second  topic that I have in my mind and maybe Stephen [8:43] if I can start with you with AI and generative AI  becoming so much easily accessible and AIdriven [8:52] fishing and deep fake like skyrocketing. So  what are some of the actionable things like [9:00] you building a product you want to make sure like  your product is also safe before you sending uh [9:04] it to the customers. So what are the measures or  actionable things you guys are thinking about to [9:09] counter those uh deep fakes and AIdriven fishing?  You know some of the alerts that we send out [9:15] certainly coming through email is hard to prove  that came from our system right that happens a [9:19] lot. SMS texting is also another thing that people  get fished on. Uh we went to Verizon, we went to [9:24] AT&T and we had it um uh verified logoed uh when  the text comes through it actually comes through [9:30] with our logo on it now. Uh that was a big deal.  Um we also do things with the the email masking [9:35] where it comes across as colorcoded, right?  So if they don't know the color scheme of the [9:38] product and what what it's talking about, then um  that helps identify if it's a valid thing or not. [9:43] That's what we're doing right now. The the other  thing with you know um phone calls and things that [9:48] you get now, I don't know if you guys have seen  that yet, where people are calling you, they sound [9:51] like real people. or not. Um, you have to wait  for the pause. I know some people have interacted [9:55] with AI enough that like this sounds generic and  usually it is, right? And so that's the way we're [9:59] doing it now with verified email, verified text  messages and then verified landline connections. [10:05] And those are also like mostly AI enabled tool,  right? Basically not a cannibalism but again [10:10] handling AI with AI. Yeah. So baselines is a big  deal, right? Like I think the thing that we we [10:18] spend the most time on with customers is saying  is this something that we'd normally see right we [10:22] we capture enough data from the customer site from  the applications from the user behavior 7 days in [10:28] you're like okay this person always does this they  get coffee at this time not we're watching cameras [10:32] or anything like that but but you have a rhythm to  the person's day and it's something that's outside [10:36] that we can flag that saying this is a risk  behavior got it right and that that's how you do [10:40] that you can do it with a seam you can do it with  a monitoring solution but when you when you do it [10:43] with AI you really get the the full data points  Right. If you're if you're looking at it from [10:47] that perspective. Got it. So maybe Terrell and and  LL come after that. So Terrell from a CISO life [10:55] now as a founder you have a much smaller team.  Obviously you cannot have an army of people to do [11:00] a lot of the things. So how are you countering  those deep fakes and AI fishing protect your [11:06] product? Ironically um it's the the same approach.  Um it's an approach that scales. It's an approach [11:13] that a lot of people don't want to hear. Um and  that's the basics like for instance um in a large [11:19] organization um that was previously at um what  we did to um kind of strengthen our our cyber uh [11:27] operations was uh help desk that fishing is a big  one in um publicly traded companies. Um but if you [11:34] have a proper segment segmentation of identities  um a pro proper taring of identity models um like [11:41] for instance a a service desk person cannot uh  change the credentials for a privilege at uh [11:46] identity um four eyes for your financial processes  like all these foundational things becomes more [11:52] and more important um so doing the basics right  um it sounds boring but it's actually very very [11:58] effective got it thank you I'll let to your  point right So doing a baseline using AI and also [12:05] understanding the behavioral pattern and and doing  some other validation and identity segmentation. [12:11] Uh what else do you guys think like is it falling  around the same lines or something else that you [12:16] guys are thinking about to counter those fishing  attempts? I think nowadays it's very difficult to [12:22] differentiate between a valid and real legit email  and a fishing email because everyone's using AI. [12:29] Well, um I I believe that the best source uh to  to know is a valid email or not um is by verifying [12:37] the source itself. And since we we empower other  uh other clients and developers to have access to [12:44] uh real time um data, search data uh at scale. uh  we are trying to leverage that by um by building [12:54] uh an AI agent that does similar to background  check when it comes to individual but it will do [13:01] uh check for the source that these emails and  these uh phone calls coming from um by leveraging [13:08] all the internet data and and then um create  a rating score for these uh these sources and [13:15] based on that rating score we can differentiate is  it a legitim or is a fishing email. Got it. Thank [13:22] you. And so Christa to the question for you will  be like these days the recon on LinkedIn and and [13:29] all of those research has become like so real.  I remember postcoid when work from home policy [13:35] came in a lot of companies got bridged okay there  is something from HR that okay you have to return [13:40] 3 days in office and [clears throat] everybody  clicks into it right like that those are like [13:44] very real and it's like playing with the recent  thing everybody's mind so what are you doing to [13:49] protect your uh you know patients and customer  and critical I mean what you are dealing with is [13:54] critical life-saving things right so yeah and so  so I'll say from a life-saving perspective um my [14:00] team has been phenomenal I can't take credit for  this Um this is a lot my architects but um we just [14:05] implemented micro segmentation in like 46 days um  from soup to nuts. That's awesome. So across and [14:11] I'm talking about granular micro segmentation. I'm  not talking about hey we took our first step and [14:15] so in healthcare to allow autonomous um robots to  to help with um you know surgeries and stuff like [14:23] that because they're more precise, right? And so  we want to allow the business and we can't be no, [14:27] right? We all said it already. um we want to  allow them to be able to use that stuff but [14:31] we need to isolate it from other things right  and isolate other things from them and biomed [14:36] devices are notoriously like horrendous when  it comes to security because the FDA cycle and [14:40] some other things so I know I see some nods high  five on that one so we are focusing very heavily [14:45] on segmentation the identity taring as well um  that was a project this year I'll tell you though [14:50] education so I I just got my certificate in cyber  psychology I found it fascinating and started to [14:56] pursue that this year how people interact with  technology not just the cyber end of it and how [15:00] that changes what they do. And so when you start  to learn how humans think, you learn how to reach [15:05] them, but you can't rely on these, especially  in healthcare because they are pulled in every [15:09] different direction. And although they they want  to um do the right thing, they don't have time [15:13] and they might miss something. And so what we're  actually focusing on is identity validation. And [15:18] you'll hear this a lot in healthcare something  like a clear or or so you know biometric is great [15:22] for physicians for example but to say like okay if  you're calling help desk you know that the health [15:27] help desk person is a help desk person that works  here and the help desk person knows that you were [15:32] the person who is supposed to be calling against  that identity and so I think we have taken a very [15:39] significant um stance on investing in our entire  identity program um especially the non-human [15:46] because it it again compliments the speed of the  business for um for AI. Yeah, that's really good. [15:52] Like I mean I heard like doing a baselining on the  behavioral pattern micro segmentation reducing the [15:57] blast radius on the network side also assuming  compromise identity based segmentation putting [16:02] more into non-human identity and understanding  building agents to detect like the original versus [16:08] fake. So uh maybe I mean I will add like two more  things that these are like holistic like what are [16:14] the things that I am looking at as my engineering  function in my current organization also like [16:20] going back to the basics email the authenticity  validation like SPF demark making P equals to [16:26] reject like making enforcing those strictly  so that people cannot do those poof domain and [16:31] domain takedowns and also looking into some of  the two other things is that pho compliant like [16:38] Okay. Spoofing resistant uh MFA right and the best  portion I loved what you said is that the training [16:45] training the help desk and making sure like when  somebody's calling and and validating uh and and [16:52] providing information just having some out of the  band validation as well right a lot of the times [16:56] if we look at the screen and technology with AI  coming in it'll be misleading can we go back to [17:02] the basics some out of the band validation with  some triggering like if somebody is triggering [17:07] suspicion and do those things. So these are also  some of the things like just to have a defense [17:12] and depth uh to do that and I don't think there  is one perfect answer. I mean in spite of doing [17:18] everything assume bridge and your cyber resilience  that time comes into picture. So yeah I want to [17:24] add one thing um to your point and uh Chris's  point. Yeah, you can't. There's an opportunity [17:30] to do everything. But what Chrissa did say which  is very powerful is she understood the why of [17:35] what they need to protect in terms of biomedical  devices and regulatory exposure and let that be [17:39] the driver which can be a catalyst for your micro  segmentation if the business knows why you're [17:44] doing it. So it's a very great starting point and  and I worked in Humana we had when I was in Humana [17:51] Humana was the fourth largest onlineies. So a lot  of the OMT devices IoT devices was there. One of [17:57] the challenge I know in healthcare is a big thing.  Even if we know that something is running out of [18:01] date and patching, you cannot do you have to go  through FDA certification process. So you have to [18:06] do like a lot of the compensating control. So it  is more difficult even to protect those things. So [18:10] hats off to you and your team like so awesome. uh  if I can change the gear and bring something which [18:16] is probably close to everybody's heart which is  claude and open AI and hugging face and anthropic [18:23] and mythos like these are in everybody's uh like  even board ask what are you doing to protect from [18:29] mythos type of things so maybe I will start with  all right so mythos type I would like advanced [18:37] modeling is helping the bad actors to find the  vulnerabilities which used to take time and also [18:43] So to layer the vulnerabilities uh in a way so  that even couple of low impact or lowrisk things [18:50] now can become a medium or even high risk. So  what are you guys thinking about to counter those [18:56] mythos type of threat if it is gone into the bad  hands? Um it's really good question. Um, methods [19:04] is very useful uh to detect vulnerabilities but  I believe that um even methas can make mistakes [19:11] when it comes to the data because if methas  has an outdated data um uh from the internet [19:17] then it will introduce hallucinations and false  positives. Um and um I believe that the the right [19:25] usage of methods should be um um not replacing the  professionals in cyber security but replacing the [19:33] competitive work that they are doing. Um I believe  that uh it will empower security professionals [19:39] uh to uh minimize the time they were taking  to detect vulnerabilities, verify them [19:45] um uh validate if these vulnerabilities are false  positives uh not false positives. Um yeah. Okay. [19:54] Okay. So Steven if I can ask the same thing right  I completely agree like data I mean you need to [19:59] know your inventory even NIST 2.0 right now talks  about govern and understanding the context and [20:05] all is very important but now if I can switch  gear into countering those AI speed threats [20:12] right the response cannot be a manual response  right okay I will get back to you in 3 weeks SLA [20:19] is no longer [laughter] right two hours uh two  hours right so it can be even minutes so when [20:25] you are countering with AI what are the tools or  technologies I mean we hear a lot of the things [20:30] right the MCPS of the world intent validation  and and and mission validation tons of things [20:37] like it's scary as well like giving the control to  AI to do a self-riting so what are the things you [20:43] were thinking about like to counter those mythos  type vulnerability threat so so myth mythos in my [20:48] opinion is a little bit more uh marketing I think  from anthropic right um as far as what it can do [20:53] what it what it can do um internally we tested it  right and we tested against our systems um it can [20:58] find vulnerabilities not saying it can't but it  doesn't have holistic controls over the network, [21:02] right? So, it identifies a machine that's at  risk. Any other tool out there, NMAP, you know, [21:06] there's plenty of tools out there that can find  that. It's the speed at which it it it um attacks, [21:11] right? Once it finds a vulnerability and then goes  to some sort of package deployment, like that's [21:15] the that's really the difference, not identifying.  So, I want to clarify that. Um what we what we've [21:20] done is we say, okay, if we give AI the same  controls that the attacker would have, which [21:24] would be if I identify a vulnerability, then take  an attack to stop that. Right? So, it's a risk [21:29] behavior, right? We identify that first. We make  a human uh interaction before, right? So we we can [21:35] we can put them in cold uh if you will or or take  them offline or or remove them from the network [21:40] or remove them from an access to an application  and then and then write something um either to an [21:46] analyst to a human to say, "Hey, check this out."  Like send it to the help desk or send it to a sock [21:50] um to take an action and and do it within minutes.  And so the issues that we found mostly for most [21:55] people is they're really good at protecting the  highle targets, right? the ones that usually go [22:00] or the like the workstations at the medical  offices or workstations at a power company [22:05] that someone said, "Oh, I'll patch those later."  Right? And that's what Methos is going to gonna [22:09] beat up. And so, you have to have something that  says everything on my network matters. It' be OT, [22:14] HVAC systems, right? The colonial pipeline started  that way. All those things matter now. And so, [22:19] because AI can look at all of those from one um u  one troth or one silo of information, you have to [22:27] treat it the same way, right? Right? And so if you  don't have a tool or something that's doing it the [22:30] same way, then you're at a disadvantage. Right?  So if they they took over the network, network is [22:34] secured, but maybe an application's not or there's  a cloud service that's not. And so that's what [22:38] Myth finds and that's what they go after. And  so if you have to be able to stay ahead of it, [22:42] we got to be more holistic and that that's the  only way to do it I think. Got it. And uh maybe [22:47] uh Terrell and then uh Christa I will come to you  u with AI like finding out the remediation like I [22:56] mean you know what needs to be patched and how or  maybe a compensating control AI can do that but [23:01] at the same time is it fully trustworthy? Will  you give full control in the organization? Let [23:06] me give you an example like company of like  my scale or even like my previous company is [23:10] JP Morgan or or any other large organization  with hundreds of firewalls. So uh will we be [23:18] comfortable giving AI the control? Okay, you make  the firewall changes which may have a huge impact [23:25] to the client if anything is done wrong. So what  do you think? So like can we or should we be or [23:31] is the world moving towards that direction that we  can we added that to our agentic agent to watch to [23:37] watch config files on hardware devices when they  make changes alert but yeah question was to him. [23:42] I'm sorry. Sorry. I was like, "No, it's all good."  Like, it's basically brainstorming problem back in [23:47] the day. But, uh, um, in general, um, when I think  of the the threats of a of a mythos style model, [23:56] I I'm not necessarily so up in arms about  the capabilities of the model itself. Um, [24:02] what I think about is the changes in the  volumes of the attack that's driven by AI. [24:06] Um that was actually a really good example that  was posted yesterday um by a new company from uh [24:13] Kevin Mandia started Mandia. Um it was actually a  control test of an offensive AI versus a defensive [24:20] AI. So the defensive AI was fully autonomous.  Offensive AI was uh fully autonomous. I think [24:25] it ran across three days like 17 million events  and 238 findings um of continuous uh attack life [24:33] cycles um mapping and and typing exploitations but  at the end of it without bound interest there was [24:40] I think 38 attack paths mapped out of 17 million  uh attacks um and uh all of them were defended [24:47] by the defensive AI. Um, so I think the problem  isn't necessarily um can we or should we um try [24:57] to defend against almost a near infinite game if  the cost of the AI is cheap. Um, so if you I have [25:07] to think of it in terms of economics and burning  tokens to continuously defend against a continuous [25:12] amount of offensive AI attacks and it seems like  that wouldn't be a very good idea because you're [25:17] going to fail first. Yeah. So either local AI  and um some sort of approach to handle internal [25:25] configuration management or an aggressive  descoping and focus on what matters the most [25:29] for your business. Um I think if you try to do  everything everywhere all at once, I think you're [25:34] going to spend a lot of money and not be able  to uh get the ROI from that investment and even [25:39] the impact of a single mistake could be even more  catastrophic than the actual risk was. Right. So [25:44] if you do that, yeah, so we'll say so and can you  imagine uh everything becoming autonomous and then [25:50] something breaking in the chain and and [clears  throat] there goes your entire security program, [25:54] right? So you take a risk based approach, high  impact, high risk, you have a human in the loop, [25:58] low impact, low risk, maybe you you automate that.  And so there's a whole um there's a whole program [26:03] you can write around like what are you going to  automate where it makes sense to your point from [26:06] a risk based decision because healthcare there  are so many things that we would love to advance [26:11] our program but the that money needs to go towards  our patient care and patient innovation right and [26:16] and we know that we're not in this to be you know  leading the path in and blazing in cyber security [26:22] and developing new tools although my team is  like awesome and I'm sure they'd love that [26:26] we we have to keep sight of the mission. Yeah,  absolutely. And maybe like just thinking loud [26:31] maybe uh we give control to AI to make changes  with some human conditions defined where human [26:38] supervision will be there. But even where we give  control to AI, there are certain conditions which [26:43] are like a hard-coded like a deterministic  code. If it is a firewall, you cannot allow [26:49] 000000000000 to everybody, right? If it is maybe  an endpoint detection and remediation tool, [26:57] then you cannot do that. So maybe those things  an organization I mean there are like top 10 can [27:03] be extremely catastrophic like you don't want to  give the keys of the kingdom maybe you hardcode [27:07] those that oh these are absolute no anything  beyond that you can do and there are certain [27:13] other conditions where it can go through human  uh supervision as well right so all uh good like [27:20] this is definitely a great conversation thank you  for sharing your perspective maybe if I can ask [27:25] uh a different question here what will Maybe  I can if I can start with you, Christa. So, [27:31] what will be one of the biggest uh blind spot for  AI in security in coming years? Oh, blind spots. [27:39] Oh my god, I wish I had time to think about that  one. We're we're I I'm still trying to hash that [27:43] out. Um and we're still trying to find that. And  and honestly, um it's hard for me to to say that [27:50] because AI has helped us so much with our threat  management. It's finding things that, you know, [27:55] our our humans don't have the cap, not the  capabilities, but they don't have the time to do, [28:00] right? And so I I just think a lot of my concern  is around where the singularity with trust is [28:08] going with security tools. I know everything that  we spoke about kind of touches on that. And so the [28:14] guy who started working on it at MIT like 65 years  ago is predicting singularity with AI in like what [28:19] 2040 and that humans are just going to trust it  with everything. And so I think the blind trust [28:24] is something and again cyber psychology right  and and humans and it all comes back to the [28:28] human condition. I love Tron by the way if anybody  catches that. [laughter] Um and and really for me [28:35] it's are we going to trust it too soon? And so  the blind spot for me is just just the level of [28:40] trust. Got it. Got it. Uh Terrell for me um it's  something we actually we model when we think about [28:49] critical uh business processes is um I think the  green path of AI and how we're leveraging it is [28:55] great but what happens when it fails like what's  the failure mode of it like um uh last week cloud [29:01] API went down how many customer sering on cloud  API now cloud API is down you don't have humans [29:08] anymore so I think the the uh biggest blind  spot is where you've successfully adopted it, [29:14] but then failure mode of the adoption and you  don't have a a manual analog function prepared to [29:21] keep the business running. Got it, Stephen? Yeah,  I think dependencies, right? One singular AI, [29:27] I think that's exactly what happens. It's API  same thing as a cloud service, right? If you [29:31] have Azure and it goes down or whatever, it's the  same thing. Um, and if you don't if you have it [29:35] all autonomous and you don't have a manual thing,  it's definitely true. Um, I think another way to [29:39] think about it also would be if um if Enthropic  or OpenAI or some of those AIs um have a problem [29:46] with their software, right? And and you've you've  given everything to them to do everything and now [29:51] there's a zero day for that. That's a big problem,  right? And so you got to make sure that whoever [29:57] you let in the door is actually doing the things  they should do for quality assurance and disaster [30:01] recovery as well as a dependent almost as if  they were working for you as an employee. Yeah, [30:07] absolutely. What's your take on it? I think the AI  has too many cons at the moment and I'm going to [30:14] mention two blinds. You build AI product, right?  You That's correct. And I'm familiar with that. [30:20] I like the honesty. Yeah, I'm I'm familiar with  that. And I used to be a penetration tester when [30:25] I was uh many years ago and uh I believe that one  of the blind spots with AI models nowadays, they [30:32] are still not that good to replace the security  professionals. For example, when it come to [30:37] penetration testing, the AI models are not really  good when it comes to blackbox testing. But when [30:42] it comes to white white box testing, code review,  uh AI are really good because it has access to the [30:47] code. But when it's come to uh blackbox, then the  AI is not that good because it's just a repetitive [30:53] work. Uh that's one of the blind spots. And the  other one is the the as I mentioned before the [31:00] quality of the data that being used to train or  being fed to the AI. If the quality of the data [31:07] is bad then the AI performance will be really  bad. If the quality of the data is high and [31:14] um really good data that is being uh used to train  the AI model then the AI model will perform way [31:20] better. Got it. And what about the providers of  those models or even Antropic or anything we can [31:28] name what about their third party and their fourth  party risk and transparency right we don't know [31:32] what sources they're using right so do you so when  you evaluate a vendor bringing in I'm just talking [31:40] about anthropic in this case it could be anything  so do you do like checks and balances and ask make [31:46] them mandatory to make it transparent and probably  that's question number one and question number two [31:51] what they don't answer and they don't share their  sources then what do you do? Yeah, unfortunately [31:56] nowadays most of uh these AI companies they  do not provide the source where they get their [32:02] data that's being used to train their AI models.  Um if we knew what is the source of their data [32:09] that being used to uh train their models then uh  might be uh it might be good for us but um but to [32:16] be honest that doesn't matter that much. Um what  matter is the quality of the model itself. Um if [32:23] they use the good data good quality high quality  data then the the model will perform uh better [32:29] than other models. So that is the main point  of the competition nowadays between all these [32:33] AI companies is where do they acquire these data  from. Got it. Got it. Uh maybe I will start again [32:40] from the left. Christa if I can start from you and  then we'll go in uh in in sequential order. So uh [32:47] we touched a lot of the things around governance  having trust making sure like human in the loop. [32:55] So when it comes to your AI speed, vulnerability  remediation and AI enabled agents and controls you [33:03] are putting in. So how do you do your governance  like the traditional governance used to be like [33:07] large organizations 100 people GRC team like those  are not feasible right now even writing policy [33:14] right previously if we think about can we write  a policy yeah I need a week and right now you go [33:19] 30 secondond co-pilot basically says okay these  are the 20 things you can do or like healthcare [33:23] where they literally have to have a policy  statement for everything that that shouldn't [33:27] even need to exist yeah it's you can't it's not  sustainable as a system yeah so so banking so I [33:32] learned this in banking I I worked for the DoD and  then um banking and then I helped build a crypto [33:37] program and then I went into healthcare and this  has been the frontier that like come to think of [33:42] it I thought crypto would be harder. So um here we  augment and banking the reason I mentioned banking [33:48] is because they kind of beat this into my head is  you have to augment everywhere that makes sense [33:53] um and risk based right I think we all said it  with um technical controls unfortunately like [33:58] so when people talk governance I think often they  think administrative governance so think about it [34:03] right you have administrative technical you know  physical controls if you get down to your core [34:07] acumen again and so I think it's really important  to look at those policies and say like okay our [34:13] policy is don't click on weird links and people  freaking do it every day, right? And so we need to [34:19] augment with technical governance. And so when you  write those like writing a framework or you take [34:24] that stuff into account, you really have to then  I I like administrative is great because there's [34:28] a liability aspect to that. I get all that and  the precedences that are being set there. But [34:32] um when it comes down to it, you really have to  think technical-minded. And although you can't [34:36] solve everything with a tool, you have to reduce  risk in the most critical areas from a governance [34:40] perspective with technical first. unfortunately  because again humans are never going to fully be [34:46] out of the loop but also they are not infallible  and then it comes into cyber psychology right I [34:50] could talk about this for weeks and you know  you could you could be tired one day or be [34:54] in a bad mood or the insider threat like the the  human threat that is malicious non-malicious and [34:59] you can't rely on humans to be consistent um and  and effective and so that's what the technology [35:04] governance control is for but for us in AI I will  tell you we have a wonderful corporate governance [35:09] structure that assesses like should we even do  this right weeds out. Does it make sense for our [35:13] business model, our culture? Um, how is it going  to impact the patient? Is the positive impact [35:18] outweigh the negative impact so much that we want  to send that then for the security governance and [35:23] risk review? So, there's a whole chain of um  assessment that that we're very strict on to [35:28] make sure that it meets the holistic need of the  organization as well. And so, governance again [35:33] um there's technical rails in there as well from  a a um clinical governance perspective. Got it. [35:39] Thank you. So uh Terrell from your point of view  when you build a product I'm pretty sure when you [35:45] go to large enterprise they ask for like a lot  of the checkbox are you sock 2 compliant are you [35:50] XYZ compliant so I am assuming you are also  using for those governance and check I mean [35:56] you use leveraging AI heavily uh to do those  certifications so do you have any concerns [36:02] sometime the reason I ask is that it's like a  chicken and egg story right you need to have [36:06] a governance and you outsource that governance  and that governance is being done by a tool and [36:12] you don't know the source of that tool where from  they are taking those LLMs and what what is their [36:17] third party and fourth party what governance they  have actually within their system to govern you so [36:23] how do you handle and do you get this questions  like maybe if I your customer I'll be asking you [36:28] those questions how do you ensure your governance  providers are properly governed yeah I mean this [36:34] is a a great question and I've lived on both sides  of it um and um from Uh uh I've built TPR programs [36:42] and built out this uh these processes before. Um  but when I get to like the the first principles of [36:48] it all, like what what am I asking and what am  I wanting from them? I don't I'm not asking or [36:53] uh I don't believe that that SOP 2 is saying  that they have governance. I it tells me that [36:58] they're attempting to do the right things. And  why why do I need that that question answered? [37:03] Because I need the I need the service that they're  providing to be reliable. Um so on the other side [37:07] of that um I put ourselves um in the position  to say here's what we're doing to protect the [37:16] services that we're offering to you. Uh here's uh  everything that we're doing around our services, [37:20] how we build them, uh anything that you need  to make a decision. Um and so much so that we [37:26] actually created our own kind of trust mechanism  that we can continuously share the resiliency of [37:31] our company to our third parties if they want to  continuously monitor what the failure risk is to [37:35] our organization. Got it. I know like we can go  on for probably hours but the time is limited. [37:42] But I want to keep like 5 minutes for open Q&A but  maybe remaining two minutes like Stephen and Allah [37:48] right if you guys can touch on with all those  things going on like skill training talent is [37:54] a big thing right and when you are building a  product you are in a startup world that's even [37:59] more challenging right you don't have like an army  of people so how are you handling and addressing [38:04] those AI training and skills issues that's a  great question uh we go to universities in the [38:08] cities that we have presence in right in Charlotte  uh Kenya in Nairobi, we go to the AI classes and [38:14] we we talk to the top two students of the class  and they offer them internships to get an idea [38:19] of what's what's emerging and then whenever new  AI comes out um we bench test it versus our lab, [38:24] right? So when um uh Deep Seek came out or Menthos  came out, we we run it against protocols to see if [38:31] it's actually a real thing or not. That's why I  say some of that I think is is is cloud is uh is [38:35] marketing. That's how we do it though. Keep  our talent that way. It also keeps our cost [38:39] down. Got it. I'm not asking the same question to  everyone all. So from your company's perspective, [38:44] how are you keeping up with the skill development  uh uh in AI space? Um we since we are the main the [38:53] the lead in the world in providing the search  data um to AI models and uh other um use cases [38:59] um we are keeping an eye on the the AI models  that are using the most accurate and the highest [39:08] quality data. Um we believe that um AI models  that being trained on the real world uh and um um [39:17] structured data scale that are the best AI models  out there and they are the most reliable nowadays. [39:24] Um no AI models nowadays can uh until now no AI  model can replace professionals but uh hopefully [39:33] in the future with um reaching more uh data at  scale accurate data high quality data we'll have [39:40] a better AI models. Got it. So uh I will keep like  5 minutes for open question and answer but if I [39:46] can wrap up the overall conversation so far what  we had number one data is the currency so we have [39:53] to make sure like we have the data cleansing and  data accuracy before automating anything through [39:58] AI. Number two, we talked about in order to  counter the pace of the attack of AI is happening, [40:05] we have to tap on AI. But that AI should not be  completely unlist. It has to have some checks and [40:10] balances. And those checks and balances could be  some manually coded thing that you can automate [40:16] everything except this. And in some areas, we need  to have a human in the loop and supervision as [40:21] well. Uh governance is another important factor.  And again the governance the classic way what we [40:27] used to do is not going to scale and in order to  scale it up in AI versus AI governance has to be [40:33] AI enable. A lot of the tools are leveraging  AI and pretty [snorts] much there is also [40:39] a risk like or blind spot is the providers  of those tooling and the brain behind those [40:45] uh solutions that we are building how transparent  they are. Are they sharing their information of [40:50] their their third party fourth party their  sources or not? Right? So having that [40:55] uh conversation and trying to drive and build  that ecosystem in the industry will definitely [41:00] uh make it uh like more uh it will improve the  overall AI security posture and the other thing [41:07] is that the fundamental things don't change like  when it comes to a deep fake or or anything like [41:13] the email anomaly detection right which is writing  on AI the basic things like the SPF demark and DKM [41:20] and all au authenticity validation doing the  micro segmentation like identity being the [41:26] parameter treating the non-human identity could  be lowering the TTL and not trusting like and and [41:32] mutual TLS the fundamentals of cyber security  still stays the same it's a matter of like [41:37] doing those more rigorously and previously we had  those things in policy maybe in the architectural [41:44] pattern somewhere but we found that those are  not implemented right now with AI coming in [41:48] making sure those are enforced in our environment  probably is making it easier here. So, and the [41:54] last thing I like Steven your point is that with  mythos and claude and everything is coming not no [42:00] need to be afraid. There are a lot of the things  okay these problems were there before it is like [42:06] going to a human scanner right when you go to a  doctor's office and it does like a full body scan [42:10] and finds like 20,000 more people scares you maybe  from doctor's lens not are important equally so [42:16] the same thing we have to apply and keep doing the  foundations well so I know we are 2 minutes and 30 [42:22] seconds so I will open it up to the audience if  you have any question that we can answer who can [42:27] help us with uh that please hello uh Chris this  is for you since I have been in healthcare for [42:33] the last 6 years um as a lead engineer it's been  very interesting I've been all over from dev sec [42:39] to what I'm doing right now uh you mentioned  something that hit home to me so I'm curious [42:43] what your thoughts are from a security with AI in  general related around policy uh most what I see [42:49] with the policy I agree with if you work for the  government you know what that means it's very slow [42:53] it's very reactive it's do what I tell you not why  um have you seen any success on being able to push [43:01] on those governing bodies that do set these  policies whether it be internal or external [43:05] like for you FDA for us would be you know Medicare  Medicaid yeah and thank you for that and God bless [43:10] I healthcare is a wild ride um coming from tax  so uh I'll tell you that one of the best things [43:16] that you can do no matter what your vertical  is in healthcare specific I did it in finance [43:20] we had other organizations um is get involved  like AMA for example if you're in banking so in [43:25] healthcare I'm part of health sector coordinating  council and we have a direct ear to um health and [43:30] human services and and we actually help steer  what is happening with HIPPA. And so, you know, [43:35] when they came out with their proposed changes  like segmentation for everybody, I'm like, okay, [43:38] that small mediumsiz office that can barely afford  someone who knows security and also does it. Um, [43:44] you're you're going to hurt healthcare, right?  Like you're already underserved and now they're [43:48] not going to have the money to survive. It's  hard enough with policy like financial policy [43:52] to survive. And so, I it's very very important.  One thing I stress holistically in my career, no [43:57] matter what vertical you're in, is to get involved  in these organizations that help steer the policy, [44:02] understand who your um who your champions are of  that and really dedicate your time like volunteer [44:08] to that because it you can get in and if you  really know what you're doing, your experience is [44:12] almost like field CESOS invaluable to these folks  um who are going through the same thing and so [44:17] share those experiences and yeah, get get with we  get with our regulators. We're in DC all the time. [44:24] Thank you. Any other follow? Oh, sorry. Thanks for  your time. Uh so really interesting conversations [44:32] and you know the hugging face CEO said that  Chinese models are closing the gap on capabilities [44:39] with the US frontier labs and my thoughts are if  the model quality becomes commoditized does the [44:48] arms race become a fight over compute telemetry  or the authority to act and you know which of [44:57] those three do you believe is actually the real  bottleneck. Who'd like to take a step? I'd say [45:03] authority to act one. Yeah, for sure. That's all  customers we interact with all the time don't even [45:08] have policy for AI, much less which AI are they  going to use. Yeah, crypto when they were doing [45:14] regulation for crypto, dude, if we waited for the  the governmental bodies to catch up like we know [45:19] would have never survived. Yeah. So, definitely  the authority. You can't wait. Sometimes you got [45:23] to ask for forgi forgiveness. Okay, we got one  more time for one more question. This kind of a [45:30] slight follow on to the hugging face question.  I just wondered what your thoughts on like [45:36] obliterated models and using them for defensive  u operations. That's actually a great question. [45:45] Anyone? Yeah, I can take that. Um I believe that  uh most of the AI models nowadays are really [45:52] good for uh defensive and code reviews but they  are not good uh um as um offensive uh usage and [46:00] um one more thing I want to clear out nowadays  is about AI. AI is not there for to replace the [46:08] professionals. uh I believe AI should be  used to replace the repetitive work only [46:15] not the professionals in cyber security. Any  other perspective anybody would like to add? [46:21] Um the one thing I I I truly believe in and  goes back to infinite versus a kind of finite [46:27] and deterministic outcome approach. If you have  a a a great harness that's focused around your [46:33] outcomes and then supplement that with AI, I  think that's like that's the million-dollar [46:38] approach right there. Yeah. I mean, having the  clarity is always important. Like 10 years ago, [46:43] like every CIO's pet project used to be, okay,  let's go to cloud and when you ask where do you [46:48] want to do it to cloud? Oh, okay. We'll figure it  out, right? So [laughter] the same thing in AI. I [46:53] mean, I see like AI at times is an overhype. Like  there are a lot of cases where RPA deterministic [46:59] Python script or even like simple scripting is way  more beneficial than over burning the calories. So [47:06] having like stay away from hype do it where it has  the real strength is the most important thing. Got