Sysdig Replaces Web Dashboards With AI Agents to Speed Cloud Defense
Ai4
Transitioning security operations to an AI-orchestrated headless model reduces incident investigation times by 73% while cutting cross-team coordination costs.
## Background
This presentation, titled "Headless Security: Cloud Defense for an AI-Speed World With Sysdig, Inc.", was delivered by Eric Carter (Product Marketing Director at Sysdig, Inc.) and published by the Ai4 channel on YouTube on August 17, 2026. The talk addresses how adversaries leverage AI for rapid reconnaissance and exploitation, proposing a "headless security" model where control loops live inside AI coding agents rather than traditional UI consoles.
## TL;DR
Eric Carter argues that traditional security dashboards have become operational bottlenecks against modern automated threats, such as Agentic Threat Actors capable of refactoring attacks in seconds. He introduces headless security—an architecture that decouples backend telemetry from the visual UI, connecting platform APIs to AI coding agents via the Model Context Protocol (MCP) and structured vendor skills. This approach shifts security practitioners from manual console operators to strategic governors, enabling customized workflows and rapid automation. Sysdig's internal study showed this model reduced vulnerability investigation time from 45 minutes across three experts down to 12 minutes for a single analyst.
## Atmosphere & style
The session carries an authoritative, visionary, and practitioner-focused tone typical of a high-level enterprise technology keynote. Carter presents complex security architectural shifts with calm clarity, balancing serious warnings about machine-speed AI threats with practical, forward-looking solutions. Demonstrations of local AI agents and custom desktop integrations give the talk a pragmatic, hands-on feel that bridges theoretical security paradigms with real-world application.
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=0i1bOMpy_js&t=4s) Rethinking Security UI in an AI-Driven Era
Eric Carter, Product Marketing Director at Sysdig, introduces the session by establishing that while vendor presentations often focus strictly on product walkthroughs, this talk will take a broader view of security operations. Addressing an audience largely composed of security practitioners, Carter highlights a shifting paradigm inspired by modern enterprise software trends. He references a recent Salesforce announcement where the traditional web interface was rendered secondary, allowing users to execute complete workflows directly within collaboration hubs like Slack without ever logging into the main console.
This same transition is now entering the cloud defense landscape through the concept of "headless security." By using AI as an orchestration layer rather than relying on static browser dashboards, practitioners can manage complex security workflows straight from their primary workspaces. Carter notes that Sysdig has driven an AI-first culture across its entire organization to explore how these models can fundamentally accelerate security engineering and threat response.
- Salesforce's "never log in again" model serves as the architectural precedent for headless security workflows.
- Headless security shifts operational control from traditional web user interfaces into conversational AI interfaces and coding environments.
[01:32](https://www.youtube.com/watch?v=0i1bOMpy_js&t=92s) The Speed Asymmetry: Agentic Threat Actors vs. Human Defenders
Carter discusses the modern threat landscape by invoking the concept of "Mythos," illustrating how attackers continuously search for and exploit vulnerabilities far faster than internal defender teams can discover and patch them. Citing data from the Zero-Day Clock, he emphasizes that the window between vulnerability disclosure, weaponization, and active system compromise has shrunk drastically. Traditional defender SLAs, which often allow days or weeks for patching, are far too generous and ineffective against automated, machine-speed adversaries.
To illustrate this asymmetry, Carter details a ransomware attack discovered by Sysdig’s threat research team that was executed entirely end-to-end by an "Agentic Threat Actor" (ATA). The attack exhibited clear signatures of autonomous AI execution, most notably by narrating its own step-by-step reasoning in plain text—much like the raw outputs observed when interacting with models like Claude, OpenAI Codex, or Claude Code.
When encountering authentication roadblocks during the breach, the ATA dynamically refactored its approach and reasoned its way around the obstacle in just 31 seconds—a turnaround time impossible for a human adversary to match. Furthermore, this attack exhibited destructive behavior: instead of exfiltrating data, it immediately destroyed the primary targeted database. In a subsequent phase observed weeks later, the ATA encrypted a private, on-premises AI model. Restoring such models requires immense time and compute resources for retraining, and standard enterprise backup procedures rarely cover AI model artifacts adequately.
- **31 Seconds:** The time it took the Agentic Threat Actor (ATA) to autonomously refactor its attack vector after hitting an initial login failure.
- ATA attacks present novel risks by targeting non-traditional enterprise assets, such as unbacked-up proprietary AI models and training pipelines.
---
[05:16](https://www.youtube.com/watch?v=0i1bOMpy_js&t=316s) Why Dashboards Are Becoming Bottlenecks
While security dashboards were once the standard for visibility, Carter argues that they have become overly rigid for modern, fast-paced technical workflows. Most practitioners already leverage AI coding agents and text prompts daily to streamline their work. The current security operational stack forces human analysts to serve as manual "glue"—context-switching constantly between Sysdig, CrowdStrike, and various SIEM consoles to reconcile disparate signals.
This manual swivel-chair analysis creates operational bottlenecks that slow down incident response. Modern workers across finance, marketing, and engineering are increasingly interacting with enterprise systems through APIs, prompts, and centralized platforms (e.g., querying Snowflake via AI integrations). Security must adapt to this same pattern, as relying on human operators to manually filter noise and prioritize alerts under tight time constraints is no longer viable.
- Human-in-the-loop context switching across disconnected security UIs is the primary driver of overly generous response SLAs.
- Security operations must transition from manual portal navigation to API-driven, prompt-based interaction models.
---
[07:42](https://www.youtube.com/watch?v=0i1bOMpy_js&t=462s) Defining Headless Security: APIs, MCP, and AI Coding Agents
Carter explains the technical mechanics of headless security as championed by Sysdig CTO Loris Degioanni. When Sysdig first enabled complete platform operation via AI coding agents like Claude Code or OpenAI Codex, it wasn't about asking security analysts to write traditional software packages; it was about generating dynamic, real-time code to solve immediate operational security problems.
Headless security relies on decoupling the underlying security data and telemetry from the visual graphical user interface (GUI). Because most modern security platforms already expose their capabilities via REST APIs, security teams historically wrote custom scripts to pull this data into unified stacks. Headless security formalizes this approach by placing the Model Context Protocol (MCP) in front of platform APIs, allowing AI agents to seamlessly orchestrate workflows across Sysdig, GitHub, Jira, EDRs, and SIEMs.
Crucial to this model are vendor-defined "skills"—specialized Markdown files that teach the AI agent the optimal strategies for accessing data, prioritizing risks, and executing workflows. Upon receiving a user prompt, the AI agent automatically ingests these Sysdig skills, granting it deep domain competency to navigate the organization's cloud and Kubernetes environments effectively.
- Headless security decouples presentation layers from backend security engines, exposing operational functions directly to AI agents via APIs and MCP.
- Security skills are structured Markdown definition files that inform AI models how to properly query data, interpret security context, and select tools.
---
[11:36](https://www.youtube.com/watch?v=0i1bOMpy_js&t=696s) Advantages of AI-Orchestrated Security Operations
Shifting to a headless security model offers four core benefits: rapid automation, infinite scale, tailored context, and personalized control. Rather than relying on human consultation for every query, analysts leverage AI engines as built-in domain experts. This enables security operations to digest the massive telemetry volume generated by modern cloud infrastructures without succumbing to alert fatigue.
Carter highlights that headless security empowers analysts to interact with security systems on their own terms—whether that means building custom charts inside Claude, creating specialized local alert loops, or triggering actions from Slack. On Sysdig’s platform, high-fidelity kernel-level telemetry and system call (`syscall`) visibility feed directly into the AI engine. Providing grounded, highly accurate runtime data at the outset minimizes LLM hallucination and reduces unnecessary token expenditure.
Furthermore, as analysts interact with the AI agent over time—accepting recommendations or specifying operational boundaries—the system learns the specific constraints of the environment. The AI adapts to the organization's unique Kubernetes architectures, cloud configurations, and compliance requirements, accelerating future decision-making while retaining human-in-the-loop approval guardrails for high-impact actions.
- **Kernel-level telemetry:** Capturing system calls (`syscalls`) supplies high-fidelity data to LLMs, reducing hallucinations and token waste.
- **Operator to Governor:** Security teams evolve from manually executing tasks in UI consoles to governing autonomous agents that operate within pre-approved boundaries.
---
[16:15](https://www.youtube.com/watch?v=0i1bOMpy_js&t=975s) Measuring Efficiency: A Case Study in Vulnerability Investigation
To quantify the impact of headless security, Carter shares an internal Sysdig study comparing traditional cross-team vulnerability investigation against an AI-assisted workflow.
In the traditional scenario, analyzing and investigating a complex vulnerability required synchronous coordination across three distinct domain experts. This cross-functional effort took 45 minutes of combined labor, creating significant operational costs and limiting total daily throughput across the security organization.
Under the headless, agentic model, a single practitioner equipped with AI skills and direct API orchestration completed the exact same investigation in just 12 minutes. The total LLM inference cost for the automated workflow was approximately $3.75 in tokens. This shift dramatically lowers the unit cost of security investigations while exponentially increasing the organization's daily handling capacity.
| Workflow Type | Required Personnel | Time to Resolve | Financial / Token Cost |
|:--- |:--- |:--- |:--- |
| **Traditional Manual Process** | 3 Domain Experts | 45 Minutes | High cross-team labor cost |
| **Headless AI-Assisted Process** | 1 Analyst + AI Agent | 12 Minutes | ~$3.75 in API tokens |
- Headless security reduced investigation time by 73% (from 45 minutes down to 12 minutes).
- Replacing multi-person manual triaging with AI orchestration allows security organizations to scale response capacity without linear headcount growth.
---
[18:15](https://www.youtube.com/watch?v=0i1bOMpy_js&t=1095s) Real-World Demonstration: Custom Interfaces Outside the UI
In the closing minutes, Carter demonstrates practical applications of headless security directly from his desktop using Claude. Rather than navigating Sysdig’s standard browser interface, he demonstrates querying the AI agent to dynamically construct a custom risk visualization map of a Kubernetes environment based on real-time Sysdig telemetry.
He also shares how he built a custom background monitoring agent on his local machine in under five minutes. By prompting the agent to poll the Sysdig runtime API every 10 minutes for new high-severity runtime threats, he configured it to issue native macOS desktop notification cards. These alerts allow him to click through directly to specific security insights, demonstrating how practitioners can easily tailor custom, lightweight security automation to fit their exact daily workflow preferences.
- AI coding agents can generate custom visualizations and ad-hoc risk dashboards on the fly without relying on pre-built UI components.
- Local background agents can be built in minutes via plain-text prompts to continuously query security APIs and trigger native OS alerts.
[19:07](https://www.youtube.com/watch?v=0i1bOMpy_js&t=1147s) Moving from Insights to Action with AI Guides
Continuing the demonstration, Eric Carter highlights how the AI-driven system generates detailed lists of issues while surprisingly maintaining interactive UI features like hover-over capabilities. Within just seven or eight minutes of interaction, a security team can gain a complete understanding of the top threats actively### Interactive Investigation and Guided Remediation
In demonstrating the power of headless security in practice, Eric Carter walks through how an AI agent goes beyond merely outputting static lists of findings. While the system can easily display summary data—complete with surprisingly rich interface elements like hover-over capabilities—the real magic happens in the interactive workflow. Within roughly seven to eight minutes, the agent gives the user a clear understanding of the top security events and risks firing across their environment. Rather than leaving the operator to figure out the next step, the system actively prompts them with choices: "do you want to deal with one or two of these? What do you want to do? Do you want
## Notable quotes
> [00:04](https://www.youtube.com/watch?v=0i1bOMpy_js&t=4s) What if the biggest obstacle to modern cloud defense isn't your tooling or your team, it's the screen in between? — Eric Carter
> [00:04](https://www.youtube.com/watch?v=0i1bOMpy_js&t=4s) Salesforce's "never log in again" model serves as the architectural precedent for headless security workflows. — Eric Carter
> [00:04](https://www.youtube.com/watch?v=0i1bOMpy_js&t=4s) Headless security shifts operational control from traditional web user interfaces into conversational AI interfaces and coding environments. — Eric Carter
> [01:32](https://www.youtube.com/watch?v=0i1bOMpy_js&t=92s) 31 Seconds: The time it took the Agentic Threat Actor (ATA) to autonomously refactor its attack vector after hitting an initial login failure. — Eric Carter
> [05:16](https://www.youtube.com/watch?v=0i1bOMpy_js&t=316s) Human-in-the-loop context switching across disconnected security UIs is the primary driver of overly generous response SLAs. — Eric Carter
> [05:16](https://www.youtube.com/watch?v=0i1bOMpy_js&t=316s) Security operations must transition from manual portal navigation to API-driven, prompt-based interaction models. — Eric Carter
> [07:42](https://www.youtube.com/watch?v=0i1bOMpy_js&t=462s) Headless security decouples presentation layers from backend security engines, exposing operational functions directly to AI agents via APIs and MCP. — Eric Carter
> [07:42](https://www.youtube.com/watch?v=0i1bOMpy_js&t=462s) Security skills are structured Markdown definition files that inform AI models how to properly query data, interpret security context, and select tools. — Eric Carter
> [11:36](https://www.youtube.com/watch?v=0i1bOMpy_js&t=696s) Operator to Governor: Security teams evolve from manually executing tasks in UI consoles to governing autonomous agents that operate within pre-approved boundaries. — Eric Carter
> [16:15](https://www.youtube.com/watch?v=0i1bOMpy_js&t=975s) Replacing multi-person manual triaging with AI orchestration allows security organizations to scale response capacity without linear headcount growth. — Eric Carter
> [18:15](https://www.youtube.com/watch?v=0i1bOMpy_js&t=1095s) AI coding agents can generate custom visualizations and ad-hoc risk dashboards on the fly without relying on pre-built UI components. — Eric Carter
> [18:15](https://www.youtube.com/watch?v=0i1bOMpy_js&t=1095s) Local background agents can be built in minutes via plain-text prompts to continuously query security APIs and trigger native OS alerts. — Eric Carter
## Mentioned
Eric Carter, Sysdig, Inc., Sysdig, Loris Degioanni, Ai4, Salesforce, Slack, Claude, Claude Code, OpenAI Codex, CrowdStrike, Snowflake, GitHub, Jira, Model Context Protocol (MCP), Kubernetes, macOS
## Video description
Eric Carter - Product Marketing Director, Sysdig, Inc.
What if the biggest obstacle to modern cloud defense isn't your tooling or your team, it's the screen in between?
As adversaries automate reconnaissance and exploitation with AI, defenders are still context-switching between consoles, manually navigating risk at human speed; a bottleneck the industry built for itself.
This talk makes the case for headless security: a model where the control loop from discovery and prioritization to remediation lives inside AI coding agents, not a traditional UI — with security capabilities exposed as APIs, guided by agent skills, and accessed via MCP so analysts can automate and personalize workflows at scale. I'll show how this shift moves the practitioner's role from operator to governor, and what organizations need to do now to keep pace with AI-powered threats.
Join this session to see what the next generation of cloud defense looks like in practice — and leave with a clear picture of where your team should be heading.
Sysdig以AI智能体取代Web仪表盘以加速云防御
将安全运营过渡至AI编排的无头模式，可在削减跨团队协调成本的同时，将事件调查时间缩短73%。
Eric Carter指出，面对能够在数秒内重构攻击的智能体威胁行为者等现代自动化威胁，传统安全仪表盘已成为运营瓶颈。他提出了无头安全架构，将后端遥测与可视化UI解耦，并通过模型上下文协议（MCP）及结构化厂商技能将平台API连接至AI编码智能体。该方法使安全从业人员从控制台的手动操作员转变为战略管理者，从而实现定制化工作流与快速自动化。Sysdig的内部研究显示，该模式将漏洞调查时间从3名专家耗时45分钟缩减至单名分析师仅需12分钟。
传统安全仪表盘面对能在数秒内重构攻击的自动化威胁已成为运营瓶颈。
无头安全架构将遥测数据与UI解耦，通过MCP协议与结构化技能将API接入AI编码智能体以赋能定制化工作流。
Sysdig内部研究表明，该模式将漏洞调查时间从3位专家协作的45分钟降至单名分析师的12分钟。
Reading summary
Sysdig Product Marketing Director Eric Carter argues that static security dashboards cannot counter autonomous, machine-speed adversaries, necessitating a shift to "headless security" where AI coding agents orchestrate telemetry and remediation through APIs and the Model Context Protocol.
Asymmetric Attack Speeds and Dashboard Bottlenecks (1:32)
Carter highlights that the window between vulnerability disclosure and weaponization has compressed significantly, making conventional enterprise patching service-level agreements (SLAs) of days or weeks ineffective. To illustrate this machine-speed threat, Carter points to a ransomware attack discovered by Sysdig that was driven end-to-end by an autonomous Agentic Threat Actor (ATA). The attacker narrated its reasoning in real-time, refactored its approach around an authentication roadblock in 31 seconds, destroyed a primary database, and later encrypted an unbacked-up proprietary on-premises AI model.
In contrast, human defenders face structural operational bottlenecks caused by traditional security dashboards. Carter argues that analysts spend excessive time as manual "glue," context-switching between tools like Sysdig, CrowdStrike, and SIEM consoles to correlate signals. Relying on manual UI navigation to filter noise under compressed timelines prevents organizations from defending against automated attacks.
Decoupling Defense via APIs and Context Protocols (7:42)
Championed by Sysdig CTO Loris Degioanni, headless security decouples backend security telemetry from visual graphical interfaces, following enterprise precedents like executing complete Salesforce workflows inside Slack. Instead of requiring human operators to navigate web portals or write complex standalone code, the architecture allows AI coding agents—such as Claude Code or OpenAI Codex—to generate dynamic scripts to address immediate security tasks.
Under this technical model, the Model Context Protocol (MCP) interfaces with existing platform REST APIs, enabling agents to coordinate workflows across Sysdig, GitHub, Jira, EDRs, and SIEMs. Vendor-provided "skills," formatted as structured Markdown definition files, supply the AI agents with domain-specific instructions on how to query telemetry, navigate Kubernetes environments, and prioritize operational risks.
High-Fidelity Telemetry and Measurable Investigation Gains (11:36)
According to Carter, headless security grounds LLM reasoning by directly feeding high-fidelity kernel-level telemetry and system calls (syscalls) into the AI model, minimizing hallucinations and token waste. As human operators interact with the agent, the system incorporates environmental constraints and organizational compliance rules, shifting the practitioner's role from a manual console operator to a governor overseeing autonomous actions within pre-approved boundaries.
To quantify efficiency improvements, Carter presented an internal Sysdig benchmark comparing vulnerability investigations. A traditional investigation required three domain experts and 45 minutes of cross-functional coordination. Using the headless model with an AI agent and API orchestration, a single analyst completed the same investigation in 12 minutes—a 73% reduction in time—at an LLM inference cost of approximately $3.75 in tokens.
Tailored Local Automation and Interactive Guided Remediation (18:15)
Demonstrating practical desktop workflows, Carter showed how AI agents can construct custom Kubernetes risk visualization maps and ad-hoc dashboards directly inside conversational tools without relying on pre-built UI components. Using plain-text prompts, he also built a local background monitoring agent in under five minutes that polled Sysdig runtime APIs every 10 minutes and pushed native desktop alert cards for high-severity threats.
During live triage, the AI-assisted system identified top security findings within seven to eight minutes while preserving interactive elements like hover-over details. Rather than leaving the practitioner to manually determine follow-up steps, the agent actively prompted the user with guided remediation options to streamline operational response directly from the conversation interface.
Carter concludes that headless security allows security teams to scale response capacity against autonomous threats without linear headcount growth. However, realizing this model requires maintaining human-in-the-loop governance for high-impact actions and updating enterprise disaster-recovery strategies to account for newly targeted assets like proprietary AI models.
## Background
该内容是由 Ai4 频道在 YouTube 平台上发布的演讲视频，主讲人为 Sysdig 产品营销总监 Eric Carter。视频聚焦于“无头安全（Headless Security）”理念，探讨了在面对 AI 驱动的秒级自动化网络威胁时，现代云防御如何打破传统控制台与仪表盘的交互瓶颈，将安全控制回路直接嵌入 AI 编码代理中，帮助企业重塑防御范式。
## TL;DR
演讲以 Sysdig 捕获的“自主代理威胁主体（ATA）”仅用 31 秒自主纠错并摧毁私有模型的实战案例切入，揭示了传统人工在碎片化仪表盘间充当“胶水”所导致的防御严重滞后。针对这一痛点，Sysdig 提出了结合 API、模型上下文协议（MCP）与 Agent Skills 的“无头安全”架构，将专业安全能力解耦并注入通用 AI 编码代理。依托底层 Linux 内核系统调用的高保真数据，该模式从源头抑制了 AI 幻觉，将原本需要 3 名专家协作 45 分钟的排查流程缩短至 1 人加 AI 在 12 分钟内闭环（耗费仅约 3.75 美元），推动安全团队从被动的“操作员”彻底演进为高效的“编排者”。
## Atmosphere & style
整场演讲风格务实、接地气且极富说服力，Carter 巧妙地通过幽默自嘲化解了参会者对“厂商推销”的戒备心理，迅速与现场安全从业者建立起深厚共鸣。从 ATA 自主推理攻击 31 秒破局的紧张警示，到现场演示 AI 按需实时渲染交互式分析图表的行云流水，叙事节奏紧凑而层次分明。现场既弥漫着对抗超高速 AI 威胁的紧迫感，又洋溢着利用新架构解放生产力的工程师乐观主义，最终落脚在“帮安全人员找回周末”的人文关怀上，兼具前瞻性与实战感染力。
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=0i1bOMpy_js&t=4s) 范式转移：无头安全（Headless Security）将安全能力从孤立的 Web 控制台解耦，直接嵌入 AI 编码智能体与对话环境中驱动自动化编排。
- [01:32](https://www.youtube.com/watch?v=0i1bOMpy_js&t=92s) 攻防速度失衡：新型自主代理威胁主体（ATA）具备自主推理重构攻击载荷的能力，仅需 31 秒即可攻破防线，远超人类防御者的响应极限。
- [01:32](https://www.youtube.com/watch?v=0i1bOMpy_js&t=92s) 新型攻击目标：AI 攻击者不仅摧毁底层数据库，更直接瞄准缺乏备份机制且重训成本极其高昂的企业私有微调 AI 模型。
- [05:11](https://www.youtube.com/watch?v=0i1bOMpy_js&t=311s) “人类胶水”困境：传统基于仪表盘的安全运维迫使分析师在多个独立工具间手动搬运上下文，迟缓的响应 SLA 已无法抵御秒级 AI 威胁。
- [07:42](https://www.youtube.com/watch?v=0i1bOMpy_js&t=462s) 无头安全架构三要素：底层能力 API 化解耦、用于标准化跨系统工具调用的 MCP（模型上下文协议），以及基于 Markdown 定义的安全代理技能（Agent Skills）。
- [11:25](https://www.youtube.com/watch?v=0i1bOMpy_js&t=685s) 打破固化 UI 与沟通瓶颈：AI 作为全天候内置专家替代传统人工问答，赋予防御者按需在 Slack 或个人面板中定制专属工作流的自由。
- [14:30](https://www.youtube.com/watch?v=0i1bOMpy_js&t=870s) 高保真内核数据基石：将基于 Linux 系统调用（Syscalls）的底层运行时事实输入 AI 引擎，能从源头有效抑制大模型幻觉并降低 Token 消耗。
- [14:30](https://www.youtube.com/watch?v=0i1bOMpy_js&t=870s) 角色演进与人机协同：防御者正从界面点击的“操作员”转变为统领 Agent 团队的“编排者”，在关键阻断与变更节点始终保持人类在回路（Human-in-the-Loop）审批。
- [17:16](https://www.youtube.com/watch?v=0i1bOMpy_js&t=1036s) 效能与成本实测突破：实测显示复杂漏洞排查从“3 位专家耗时 45 分钟”大幅缩短至“1 人配合 AI 耗时 12 分钟且 Token 成本仅约 $3.75”，显著降低边际成本并提升响应并发度。
- [18:18](https://www.youtube.com/watch?v=0i1bOMpy_js&t=1098s) 端到端向导式落地：AI 智能体能够实现从实时风险分析、定制视图渲染到自动提 PR 与云配置修复的完整闭环，企业可从 CVE 漏洞分流等高频场景逐步切入。
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=0i1bOMpy_js&t=4s) 引言：从传统 UI 到“无头安全”（Headless Security）的范式转移
演讲者 Eric Carter（Sysdig 产品营销总监）开场向现场参会者致意，并敏锐地察觉到大家通常对“厂商宣讲”（Vendor Session）抱有戒心，担心沦为单纯的产品推销。他承诺本次演讲将跳出具体产品，聚焦于整个安全行业面临的结构性变革。在现场互动中，大部分观众都表示自己从事安全相关工作。Carter 借用 Salesforce 之前的创新理念做类比——Salesforce 曾提出“用户无需再单独登录控制台”，只要在 Slack 或日常办公界面即可完成所有业务操作。Sysdig 正在将完全相同的理念引入安全领域，即“无头安全”（Headless Security）：利用人工智能作为统一编排和执行安全工作流的中枢，摆脱传统界面的束缚。
Carter 坦言，作为产品营销人员，他深知参会者对这类头衔的固有偏见，但强调 Sysdig 全员早已深度投入到 AI 技术的实战落地中。无论是利用 Claude 还是其他前沿工具，AI 正在重塑企业防御的底层逻辑。安全团队过去习惯了登录一个个孤立的 Web 控制台，而无头安全的核心就是把安全能力直接嵌入到分析师与开发者日常使用的 AI 代理与对话环境中，实现从被动查看向主动编排的演进。
### [01:32](https://www.youtube.com/watch?v=0i1bOMpy_js&t=92s) AI 驱动的自主威胁：31秒破局与模型勒索实战案例
当前防御者面临的最严峻挑战在于攻防不对称性的急剧恶化。Carter 提到行业内讨论的“零日时钟”（Zero Day Clock）概念以及类似 Mythos 的自动化攻击能力——攻击者发现漏洞和武器化利用的速度正在呈指数级提升。为了应对这种“以火攻火”（Fight fire with fire）的局面，防御方必须直面由 Sysdig 威胁研究团队捕获的全新威胁形态：“自主代理威胁主体”（Agentic Threat Actor，简称 ATA）。这是一种完全由 AI 端到端自主运行的勒索攻击，性质与此前针对 OpenAI 和 Hugging Face 生态出现的自动化威胁类似。
研究团队通过多项关键特征确认了该攻击完全由 AI 自主驱动，其中最典型的就是攻击程序在执行过程中会“实时自我叙述推理逻辑”（Narrated its own reasoning），像人类使用 Claude Code 或 OpenAI Codex 一样在文本中输出“此路不通，接下来尝试方案 B”的思考过程。更令人震惊的是其自主纠错能力：当攻击程序因语法或机制错误遭遇登录失败时，它并没有放弃，而是自主推理重构（Refactor）了攻击载荷，整个过程仅耗时 31 秒便成功攻破防线。这种响应速度远超人类防御者的应对极限。
在实际破坏阶段，该 ATA 展现出了极高的破坏性和残酷性：
- **第一阶段（数据库毁灭）**：自称勒索软件并留下了比特币勒索地址，但实际上根本没有外发窃取数据，而是直接加密并摧毁了底层数据库。
- **第二阶段（私有 AI 模型摧毁）**：攻击者在数周后瞄准了受害者内部经过大量时间与算力微调训练的私有 AI 大模型，并将其直接加密摧毁。
- **备份盲区**：传统运维体系往往缺乏针对私有 AI 模型的备份机制，重新训练成本极其高昂；由于攻击全程在极短时间内自主闭环，当安全团队在告警系统中看到异常时，“攻击早已结束”（The attack's over）。
### [05:11](https://www.youtube.com/watch?v=0i1bOMpy_js&t=311s) 仪表盘困境与人类胶水瓶颈：滞后的安全运营现状
传统安全运维模式严重依赖各类可视化仪表盘（Dashboards）。尽管仪表盘在呈现信号和统计图表方面依然有用，但其固化、死板的交互形式已无法适应现代高节奏的工作方式。现场几乎所有参会者都在日常工作中频繁使用 AI 编码代理或文本提示词来定制自己的工作流。然而，安全团队的现状却是被割裂在多个控制台之间——从 Sysdig、CrowdStrike 到 SIEM 系统，分析师不得不充当不同工具之间的“人类胶水”（The glue between all these things），手动搬运上下文与数据。
```mermaid
flowchart TD
accTitle: 传统安全运营的人类胶水瓶颈与无头安全编排对比
subgraph Legacy ["传统模式：分析师充当人类胶水 (低效且延迟长)"]
D1["Sysdig 控制台"] --> H["安全分析师<br/>(手动切换上下文/过滤告警)"]
D2["CrowdStrike 仪表盘"] --> H
D3["SIEM / 日志平台"] --> H
H --> ACT["人工响应决策<br/>(SLA 过于宽松/滞后)"]
end
subgraph Headless ["无头安全模式：AI 代理 + MCP 统一编排 (即时响应)"]
API1["Sysdig API"] --> MCP["MCP 服务层"]
API2["CrowdStrike / SIEM"] --> MCP
API3["GitHub / Jira"] --> MCP
MCP --> AGENT["AI 编码代理<br/>(Claude Code / Codex)"]
SKILL["Sysdig Agent Skills<br/>(最佳实践 Markdown)"] --> AGENT
AGENT --> EXEC["自动化/个性化工作流执行"]
end
```
这种碎片化的操作模式导致企业的安全服务等级协议（SLA）显得过于宽松和迟缓，根本无法抵御以秒级速度推进的 AI 威胁。在非安全部门（如市场、财务），员工已经习惯通过 API 将 AI 接入 Snowflake 等数据底座直接生成洞察，而安全领域却因责任归属顾虑（“出事了该解雇谁？”）仍然坚持让全人工去完成海量告警的优先级排序（Prioritization）。但在攻击速度逼近即时化的今天，人工逐一排查告警的时间窗口早已荡然无存。
### [07:42](https://www.youtube.com/watch?v=0i1bOMpy_js&t=462s) 什么是无头安全：API、MCP 与 Agent Skills 的协同编排
针对这一困局，Sysdig 提出了基于 AI 编码代理（AI Coding Agents）的无头安全范式。当 Sysdig 首席技术官 Loris Degioanni 最初提出“让安全团队在编码代理中操作所有 Sysdig 功能”时，Carter 也曾疑惑：“安全人员又不是写商业软件的开发者，为什么要用 Coding Agent？”但随后团队意识到，这并非让安全人员去开发对外售卖的软件包，而是利用代码生成与执行能力，在 Claude Code 或 Codex 中即时生成、编排符合当下上下文的安全工作流。
无头安全架构由以下核心要素构成：
- **UI 与数据解耦**：保留底层强大的威胁检测与分析能力，但将其全部通过标准 API 暴露，拥抱用户脱离控制台、通过脚本聚合跨平台数据的真实习惯。
- **模型上下文协议（MCP）**：在 API 前端部署 MCP（Model Context Protocol）服务，使 AI 能够干净、标准地同时调用 Sysdig、EDR、SIEM、Jira 以及 GitHub 等全链路工具。
- **安全代理技能（Agent Skills）**：作为整个体系的关键组件，厂商将自身深厚的安全知识与最佳实践固化为结构化的技能文件。这些技能本质上是简单清晰的 Markdown 文件，明确定义了数据访问路径、分析逻辑以及推荐 AI 调用的辅助工具。
当分析师向 AI 提出安全排查需求时，AI 代理首先加载 Sysdig Skills，瞬间具备专业级的安全工具调用与威胁研判能力。这使得通用 AI 代理转变为深谙企业安全环境的专家级助手，自主完成跨系统的数据调取、降噪排序与工单联动，将安全运维人员从繁重的界面点击中彻底解放出来。
- 很多客户此前会专门向 Sysdig “道歉”，表示虽然产品 UI 做得很棒，但他们为了跨工具流转，往往直接写脚本调用底层 API；无头安全正是对这一真实需求的顺势拥抱。
- Agent Skills 的定义门槛极低，任何人都可以通过简单的 Markdown 编写并复用针对特定安全场景的操作流。
### [11:25](https://www.youtube.com/watch?v=0i1bOMpy_js&t=685s) 无头安全的核心价值：自动化、规模化与完全定制
向无头安全模式演进带来了几个显而易见的巨大优势，首当其冲的便是自动化与提速。在现代安全运营中，安全团队长期受制于繁杂的人工沟通壁垒——当分析师遇到不理解的告警时，往往不得不转头去问同事：“哥们，这是什么意思？答案是什么？”这种依赖人工答疑的模式极大地拖慢了响应节奏。而当 AI 引擎内嵌了专业安全能力后，这种内部摩擦被大幅消除。面对日常海量涌入的告警与数据洪流，单纯依靠人力根本无法看清全局，唯有为 AI 提供精准的数据与充分的上下文，安全团队才能真正具备规模化处理威胁的能力，以最高效的路径达成防御目标。
另一个根本性的变革在于工作流的彻底定制化。传统安全产品往往将厂商自认为“正确”的工作流强行固化在 UI 界面中，迫使企业用户去适应繁琐的界面逻辑。而在无头架构下，控制权回到了防御者手中。分析师可以根据个人偏好与团队习惯，自由构建专属于自己的图表、交互界面，或是直接将自动化工作流无缝接入 Slack 等协同工具，彻底告别千篇一律的控制台切换。
- **消除沟通瓶颈**：AI 作为全天候内置专家，替代传统“转头问同事”的人工知识检索模式。
- **上下文驱动的规模化**：在海量数据冲击下，依靠精准上下文实现从发现风险到达成目标的规模化处置。
- **自定义界面与工作流**：打破固化 UI 限制，允许分析师按需在 Slack 或个性化面板中驱动安全操作。
### [13:00](https://www.youtube.com/watch?v=0i1bOMpy_js&t=780s) Sysdig 的实践：通过技能与 MCP 赋能 AI 编码智能体
Sysdig 正在将其积累的底层安全能力全面解耦并对外开放，使开发者的 AI 编码智能体（AI Coding Agent）能够直接调用安全洞察。无论是在编写代码时需要评估安全隐患、推荐漏洞修复方案，还是寻找阻断攻击路径的最佳手段，AI 智能体都可以通过特定技能（Skills）以及模型上下文协议（Model Context Protocol, MCP）与 Sysdig 平台无缝交互。
为了说明这种个性化智能体有多容易构建，演讲者分享了自己在桌面端的一个实际体验：他希望构建一个每 10 分钟自动轮询一次高危运行时威胁的专属 Agent。起初他要求 Agent 发送邮件提醒，但 Agent 主动建议：“发邮件并不是好方案，不如在你的 Mac 桌面上弹出系统通知如何？”在采纳建议后，短短五分钟内他就获得了一个完全按个人习惯运行的自动化工具，一旦捕获到运行时威胁就会弹出原生通知，点击即可直接跳转至详细洞察。这种由个人随手创建、贴合个人工作流的智能体协作，彻底颠覆了传统的监控方式。
```mermaid
flowchart LR
accTitle: Sysdig MCP 赋能 AI 智能体工作流架构
subgraph SysdigCore["Sysdig 底层平台"]
A["内核级 Syscall 数据"] --> B["高保真威胁引擎"]
end
subgraph InterfaceLayer["集成与交互层"]
B --> C["MCP (模型上下文协议)"]
B --> D["安全 Agent Skills"]
end
subgraph AgentLayer["AI 智能体与客户端"]
C --> E["AI 编码智能体 / LLM"]
D --> E
E --> F["本地桌面通知 / Slack"]
E --> G["自动生成 PR / 配置修复"]
end
```
### [14:30](https://www.youtube.com/watch?v=0i1bOMpy_js&t=870s) 高保真内核数据与人机协同：从“操作员”演进为“编排者”
AI 系统的输出质量完全取决于输入数据的精度。为了避免大语言模型在安全分析中出现“幻觉”、得出错误结论甚至浪费大量 Token，Sysdig 将底层基于 Linux 内核系统调用（Syscalls）捕获的高保真运行时数据直接喂给 AI 引擎。拥有了高保真事实作为基石，AI 才能摆脱无休止的猜测，做出精准判断。
在此基础上，AI 智能体与安全人员之间形成了一种动态学习机制。在日常的人机交互与问答协作中，智能体会逐渐熟悉当前环境的运行规则以及团队的处置偏好——记住哪些告警可以直接点击“接受”、哪些属于误报、当前的 Kubernetes 集群和云环境有何特异性。随着磨合加深，智能体能够以更快的速度按照团队习惯生成报告并推导结论。但在关键时刻，安全人员依然保有绝对控制权，必须设置“暂停键”由人类完成最终审批确认，从而在安全交互中实现持续自我进化。
这种演进彻底改变了安全团队的角色定位：安全人员不再需要每天紧盯仪表盘、疯狂点击界面或是四处求助，企业也不再需要仅仅为了处理更多告警而无限制增加编制。防御者正在从琐碎任务的“操作员（Operator）”转变为宏观调度的“编排者（Orchestrator）”，凭借完整的审计追踪机制统领智能体团队协同作战。
- **高保真内核数据**：基于底层系统调用（Syscalls）提供事实依据，从源头抑制 AI 幻觉并降低 Token 消耗。
- **自适应环境学习**：智能体在日常协作中学习企业独特的 K8s/云环境配置与审批偏好。
- **人类在回路（Human-in-the-Loop）**：关键变更必须经过人类确认与审计，确保安全可控。
### [17:16](https://www.youtube.com/watch?v=0i1bOMpy_js&t=1036s) 效率与成本对比实测：45 分钟跨部门协同 vs 12 分钟单人加 AI
为了量化无头安全与 AI 协作的实际价值，Sysdig 在内部进行了一场真实的漏洞排查对比测试。在传统模式下，解决一个复杂的漏洞调查需要调动三位来自不同领域的资深专家共同协作。尽管沟通已经非常高效，三人通力合作依然耗费了整整 45 分钟才完成全流程；折算三名工程师的工时薪资后，单次处置的成本相当高昂，这也直接锁死了企业每天能够处理的漏洞上限。
而在 AI 辅助模式下，仅需一名普通分析师在 AI 智能体的协助下提问与排查。即使该分析师不具备跨领域专家的全面知识，AI 也能快速补齐背景并给出指引，整个调查在 12 分钟内便彻底完成，消耗的 Token 成本仅为约 3.75 美元。这不仅使日常漏洞处置吞吐量实现了数倍增长，还将单次响应的边际成本压低到了极具性价比的区间。
| 对比维度 | 传统人工协同模式 | AI 智能体辅助模式 |
|:--- |:--- |:--- |
| **参与人员** | 3 位跨领域领域专家 | 1 位普通分析人员 + AI 智能体 |
| **排查耗时** | 45 分钟 | 12 分钟 |
| **直接成本** | 3 名资深工程师工时费用 | 约 $3.75 美元 Token 费用 |
| **业务吞吐** | 受限于专家人数与精力，容量瓶颈严重 | 处理能力大幅跃升，可实现高并发响应 |
### [18:18](https://www.youtube.com/watch?v=0i1bOMpy_js&t=1098s) 现场演示与落地路径：从 CVE 分流到“找回你的周末”
在最后的演示环节中，演讲者展示了利用 Claude 结合 Sysdig 数据直接生成按需定制界面的能力。分析师只需用自然语言要求 Claude 分析 Kubernetes 环境中的最高风险区域，AI 便会实时拉取 Sysdig 数据并渲染出一张具备交互悬停能力的定制图表，而这种视图在传统的 Sysdig 官方 UI 中根本不复存在。类似地，分析师可以在七八分钟内理清环境中正在触发的所有主要威胁，并由智能体主动引导推进后续动作：“你打算处置其中的哪一项？需要立即采取行动还是进一步深度调查？”
```mermaid
flowchart TD
accTitle: 从威胁发现到修复的 AI 引导闭环
A["自然语言提问 / 事件触发"] --> B["AI 实时拉取 Sysdig 高保真数据"]
B --> C["动态生成交互式风险视图"]
C --> D["智能体主动推荐处置策略"]
D --> E{"人类审批确认"}
E -- "批准" --> F["执行 AWS/K8s 自动化配置变更或自动提 PR"]
E -- "拒绝/调整" --> G["调整策略并记录审计日志"]
```
演讲者强调，这种模式的精妙之处在于它“不是一个简单的聊天机器人（Chatbot），而是一个端到端的向导（Guide）”，能够带领分析师完整走通风险发现、研判、自动生成 Pull Request 以及云端配置修复（例如直接在 AWS 中执行经过人类批准的安全变更）的全流程。
对于企业而言，迈向无头安全不必追求“一步到位”，完全可以从痛点最明显的单点切入：
- **从高频场景起步**：将 CVE 漏洞分流（Triage）与自动化修复作为切入点。
- **打通研发与协同工具**：无缝对接 Jira 工单系统，自动生成代码修复 PR。
- **工具生态全面编排**：打破厂商壁垒，将企业现有的所有安全工具编排为统一的智能防御矩阵。
演讲最后总结道，推动这一变革的根本目的，是让安全工具真正以编排化、自动化的方式为防御者服务，把安全人员从无休止的告警与加班中解脱出来，“让你重新找回属于自己的周末”。
## Notable quotes
> [00:04](https://www.youtube.com/watch?v=0i1bOMpy_js&t=4s) Salesforce 曾提出用户无需再单独登录控制台，只要在 Slack 或日常办公界面即可完成所有业务操作，Sysdig 正在将完全相同的理念引入安全领域。 — Eric Carter
> [01:32](https://www.youtube.com/watch?v=0i1bOMpy_js&t=92s) 攻击程序在执行过程中会实时自我叙述推理逻辑，像人类使用 Claude Code 或 OpenAI Codex 一样在文本中输出‘此路不通，接下来尝试方案 B’的思考过程。 — Eric Carter
> [01:32](https://www.youtube.com/watch?v=0i1bOMpy_js&t=92s) 当安全团队在告警系统中看到异常时，攻击早已结束（The attack's over）。 — Eric Carter
> [05:11](https://www.youtube.com/watch?v=0i1bOMpy_js&t=311s) 安全团队的现状是被割裂在多个控制台之间，分析师不得不充当不同工具之间的人类胶水（The glue between all these things）。 — Eric Carter
> [07:42](https://www.youtube.com/watch?v=0i1bOMpy_js&t=462s) 很多客户此前会专门向 Sysdig 道歉，表示虽然产品 UI 做得很棒，但他们为了跨工具流转，往往直接写脚本调用底层 API；无头安全正是对这一真实需求的顺势拥抱。 — Eric Carter
> [07:42](https://www.youtube.com/watch?v=0i1bOMpy_js&t=462s) 安全人员又不是写商业软件的开发者，为什么要用 Coding Agent？ — Eric Carter
> [11:25](https://www.youtube.com/watch?v=0i1bOMpy_js&t=685s) 哥们，这是什么意思？答案是什么？ — Eric Carter
> [13:00](https://www.youtube.com/watch?v=0i1bOMpy_js&t=780s) 发邮件并不是好方案，不如在你的 Mac 桌面上弹出系统通知如何？ — Eric Carter
> [14:30](https://www.youtube.com/watch?v=0i1bOMpy_js&t=870s) 防御者正在从琐碎任务的操作员（Operator）转变为宏观调度的编排者（Orchestrator），凭借完整的审计追踪机制统领智能体团队协同作战。 — Eric Carter
> [18:18](https://www.youtube.com/watch?v=0i1bOMpy_js&t=1098s) 这种模式的精妙之处在于它不是一个简单的聊天机器人（Chatbot），而是一个端到端的向导（Guide）。 — Eric Carter
> [18:18](https://www.youtube.com/watch?v=0i1bOMpy_js&t=1098s) 推动这一变革的根本目的，是让安全工具真正以编排化、自动化的方式为防御者服务，让你重新找回属于自己的周末。 — Eric Carter
## Mentioned
Eric Carter, Sysdig, Loris Degioanni, Salesforce, Slack, Claude, Claude Code, OpenAI, OpenAI Codex, Hugging Face, Mythos, CrowdStrike, SIEM, Snowflake, Jira, GitHub, Model Context Protocol (MCP), AWS, Kubernetes, Linux
## Video description
Eric Carter - Product Marketing Director, Sysdig, Inc.
What if the biggest obstacle to modern cloud defense isn't your tooling or your team, it's the screen in between?
As adversaries automate reconnaissance and exploitation with AI, defenders are still context-switching between consoles, manually navigating risk at human speed; a bottleneck the industry built for itself.
This talk makes the case for headless security: a model where the control loop from discovery and prioritization to remediation lives inside AI coding agents, not a traditional UI — with security capabilities exposed as APIs, guided by agent skills, and accessed via MCP so analysts can automate and personalize workflows at scale. I'll show how this shift moves the practitioner's role from operator to governor, and what organizations need to do now to keep pace with AI-powered threats.
Join this session to see what the next generation of cloud defense looks like in practice — and leave with a clear picture of where your team should be heading.
[Source](https://www.youtube.com/watch?v=0i1bOMpy_js)
Source quotes
[0:04] Eric Carter: Salesforce 曾提出用户无需再单独登录控制台，只要在 Slack 或日常办公界面即可完成所有业务操作，Sysdig 正在将完全相同的理念引入安全领域。
[1:32] Eric Carter: 攻击程序在执行过程中会实时自我叙述推理逻辑，像人类使用 Claude Code 或 OpenAI Codex 一样在文本中输出‘此路不通，接下来尝试方案 B’的思考过程。
[1:32] Eric Carter: 当安全团队在告警系统中看到异常时，攻击早已结束（The attack's over）。
[5:11] Eric Carter: 安全团队的现状是被割裂在多个控制台之间，分析师不得不充当不同工具之间的人类胶水（The glue between all these things）。
[7:42] Eric Carter: 很多客户此前会专门向 Sysdig 道歉，表示虽然产品 UI 做得很棒，但他们为了跨工具流转，往往直接写脚本调用底层 API；无头安全正是对这一真实需求的顺势拥抱。
[7:42] Eric Carter: 安全人员又不是写商业软件的开发者，为什么要用 Coding Agent？
[11:25] Eric Carter: 哥们，这是什么意思？答案是什么？
[13:00] Eric Carter: 发邮件并不是好方案，不如在你的 Mac 桌面上弹出系统通知如何？
[14:30] Eric Carter: 防御者正在从琐碎任务的操作员（Operator）转变为宏观调度的编排者（Orchestrator），凭借完整的审计追踪机制统领智能体团队协同作战。
[18:18] Eric Carter: 这种模式的精妙之处在于它不是一个简单的聊天机器人（Chatbot），而是一个端到端的向导（Guide）。
[18:18] Eric Carter: 推动这一变革的根本目的，是让安全工具真正以编排化、自动化的方式为防御者服务，让你重新找回属于自己的周末。
Source quotes
[0:04] Eric Carter: What if the biggest obstacle to modern cloud defense isn't your tooling or your team, it's the screen in between?
[0:04] Eric Carter: Salesforce's "never log in again" model serves as the architectural precedent for headless security workflows.
[0:04] Eric Carter: Headless security shifts operational control from traditional web user interfaces into conversational AI interfaces and coding environments.
[1:32] Eric Carter: 31 Seconds: The time it took the Agentic Threat Actor (ATA) to autonomously refactor its attack vector after hitting an initial login failure.
[5:16] Eric Carter: Human-in-the-loop context switching across disconnected security UIs is the primary driver of overly generous response SLAs.
[5:16] Eric Carter: Security operations must transition from manual portal navigation to API-driven, prompt-based interaction models.
[7:42] Eric Carter: Headless security decouples presentation layers from backend security engines, exposing operational functions directly to AI agents via APIs and MCP.
[7:42] Eric Carter: Security skills are structured Markdown definition files that inform AI models how to properly query data, interpret security context, and select tools.
[11:36] Eric Carter: Operator to Governor: Security teams evolve from manually executing tasks in UI consoles to governing autonomous agents that operate within pre-approved boundaries.
[16:15] Eric Carter: Replacing multi-person manual triaging with AI orchestration allows security organizations to scale response capacity without linear headcount growth.
[18:15] Eric Carter: AI coding agents can generate custom visualizations and ad-hoc risk dashboards on the fly without relying on pre-built UI components.
[18:15] Eric Carter: Local background agents can be built in minutes via plain-text prompts to continuously query security APIs and trigger native OS alerts.
Original transcript
[0:04] Hey, thanks for taking the
time to come. I know this is a
[0:08] vendor session which sometimes you're like,
hey, what am I going to learn about my product?
[0:13] But I'll try to keep it more general than
that. How many folks here in the security
[0:17] space in some way, shape, or form?
Yeah, a lot of you. Great. Great. Um,
[0:24] and that these days means a lot of us, right? So,
one of the things I wanted to do today is just
[0:29] introduce this concept of headless security.
How many folks uh saw the the the Salesforce
[0:34] announcement a few couple months ago where they're
like, "Hey, you never have to log in again."
[0:39] Yeah. Anyone seen that? Like we just get it from
Slack or wherever you're working from. You can now
[0:45] do all of the things you would have done had
you logged into Salesforce. And we're this
[0:49] is a similar concept in that we're moving this
into the security space leveraging AI as a place
[0:56] where we can start to operate or orchestrate our
security workflows. So let's take a look at uh
[1:02] what I've prepared for today and if there's
time I'm going to show you my view of
[1:08] what this looks like straight out of claude. Now
[snorts] I am just to be full disclosure I'm uh on
[1:14] the product marketing team at Cyig. That's kind of
a bad title for sometimes from these conferences,
[1:19] but don't worry, I'm deep into it and we've
gotten everyone in the company deep into AI and
[1:24] how it's going to help us. So, what problem are we
trying to solve? Now, if I said the word mythos,
[1:32] what's the first thing that comes to mind, right?
Yeah, it's kind of a single word now for invoking
[1:39] this ah crap that's the thing that's going to find
all my vulner vulnerabilities faster than I am
[1:44] going to exploit them faster than my team and
it's going to put me in a lot in a world of hurt
[1:49] right so you can think about this two ways one
uhoh we'll never be able to keep up but instead of
[1:57] thinking that way what can we do to try and match
pound-for-pound or you know fight fire with fire
[2:03] choose your cliche and it really is a a look at
this comes from something called zero day clock.
[2:09] Attackers are doing things. It doesn't take very
long when they discover stuff. It doesn't take
[2:14] long to to weaponize it and actually get make
have their way with your systems, your data.
[2:20] And the defenders in many shops are
still dealing with SLAs's that are far
[2:25] too let's call it uh generous far too
generous for us to actually make an impact
[2:31] on these kind of exploits on these kind of
attacks. So something that we're trying to help
[2:38] the world solve help our customers solve let
me give you an example of what we're facing.
[2:44] This is not unlike the open AI/hugging face
thing that happened just a couple of weeks ago.
[2:51] And what I mean by that is consider that we
now have these things. We're calling them
[2:56] uh agentic threat actors or ATA because we we
all love acronyms. Um and it is a ransomware
[3:05] attack that happened that our threat
research team found that's run entirely
[3:12] end to end by AI. Now, how did we know
that? There were lots of telltale signs
[3:16] that uh there wasn't really a human directing
everything, right? It one I the one thing I point
[3:22] to is it narrated its own reasoning. So, you
can actually see along the way kind of the the
[3:27] plain text of it saying, "Okay, well, this didn't
work, so I'm going to try that. This is, you know,
[3:32] what we see when we're interacting with whatever
it might be, Claude, Cloud Code, OpenAI, Codeex,
[3:37] and whatnot." So very clear that AI was kind of
running the show here. Now the big thing that
[3:45] is a bit troubling is when it ran
into roadblocks, what did it do?
[3:51] It reasoned its way around it. So it took
31 seconds when it could not log in because
[3:56] it was using some sort of whatever it was uh
syntax or the mechanism was incorrect. So it
[4:02] refactored and tried again. 31 seconds later it
was in. So that's pretty crazy. It takes just a
[4:08] human is going to take longer than that. I might
be able to figure it out, but it's going to be
[4:12] longer than 31 seconds. And so, we're not going to
be able to see things most of the time in in time.
[4:18] We're not going to be able to respond in time.
And what was particularly cruel about this one,
[4:22] it said it was a ransomware attack, so it was
encrypting everything. Even gave a Bitcoin address
[4:27] and so on and so forth. [snorts] The problem was
it never actually took any of the data out. So,
[4:33] it basically destroyed pass one. It targeted
a database, destroyed it. There was no way
[4:38] to get it back. Well, if the customer has
backups, okay, we might be in good shape.
[4:42] Phase two, which we saw a couple weeks later,
it actually destroyed encrypted a an a AI model
[4:50] private on. And the problem with that is it takes
a lot of time, training, and so on and so forth
[4:56] for us to actually make those things the way
we want them. And if someone destroys it,
[5:00] the problem becomes that's an expensive
proposition to redo all that work. And most
[5:06] people are not doing your traditional backups
on AI models. So by the time we see things,
[5:11] the attack's over. That's the point. Now, we
all love our dashboards or we used to anyway.
[5:16] Um, but increasingly that becomes while it's great
for seeing the problems and the signals and so on,
[5:24] it's a bit too rigid for how we're working
today. How many folks are actively using um
[5:31] AI coding agent AI text prompt every day in your
job to some degree? Yeah, most of you. Right. So,
[5:42] I'm kind of jumping ahead of myself, but the
problem that I see or that we see is that
[5:47] we're starting to use these tools to customize
how we do work to work where I want to work. I
[5:53] brought up the Salesforce thing because a lot of
people are just sitting in Slack doing whatever
[5:57] they do and want to be able to get insights and
information and maybe even trigger something from
[6:01] there before maybe they jump out and do something
else. In our current world where I have a cystic
[6:06] dashboard, I've got a crowd strike dashboard.
I've got my SIM dashboard. I'm sort of the glue
[6:12] between all these things or even different teams
are the glue between all these things. And so
[6:18] that's where the risks start to pile up. That's
where I said that our SLAs's are a bit too
[6:23] generous today in our current security teams and
world. And that's changing quickly. So, this is
[6:31] kind of where I was headed with this with AI.
What are we doing today? We're asking questions.
[6:37] We're building and triggering workflows in lots
of different places. Not even just take security
[6:42] off the table for a second. just in general
in in in a marketing job in a finance job
[6:48] in lots of jobs across an organization and we're
working through promps APIs it's systems around
[6:54] us right for instance at cyig we've plugged in
um our AI cloud stuff into snowflake right and
[7:00] other systems that we have so now we can run these
reports and most people are getting used to doing
[7:05] that work that way but what about security is
still catching up now [snorts] it assumes a lot
[7:12] of times is that we're going to decide because
we're worried. Who do you fire when something
[7:17] goes wrong? So, we still expect a human to decide
what matters and what to do next. Prioritization.
[7:23] Now, prioritization, filtering through the
noise is a big huge challenge for teams.
[7:29] Now, that's why we're getting paid, but there's a
lot of work we'd rather be doing. So [snorts] when
[7:34] we have time which we don't have this
is all fine but we again we don't have
[7:42] time anymore. So what do we see happening? We
see AI coding agents and I say that specifically
[7:48] AI coding agents and this thing we call headless
security coming into play. The first time
[7:55] our CTO Loris said yeah we're we're doing this
integration. We're making it so that everything
[8:01] cystic can be operated out of a coding agent.
So security teams can do this. And I thought
[8:07] again my limited mind it was like coding but
they're security. They're not developers.
[8:13] They're co coding. Well this isn't about
developing a software package and setting it
[8:17] aside and selling it to the market. This is about
getting in real time workflows that work for me
[8:23] for solving my security problems. It just so
happens that it's clawed code or it's codecs.
[8:29] And yes, there's code being generated behind
the scenes, but that's what's getting us faster
[8:34] into some of these challenges and problems
that we have out there trying to keep up
[8:39] with what's happening in the security space. So,
when we look at what we mean by headless security,
[8:44] it's a few things. One, very clearly, it's like
I can still have my UI, but I'm going to decouple
[8:51] this from the data itself. So all those insights
that we deliver or whatever tools are in your shop
[8:57] deliver, they're usually already available by
APIs. And a lot of folks were already doing this
[9:02] with scripts and things like that. How many people
have ever built like a script that pulls out of
[9:06] an API? Yeah. Instead of using the UI, it's like
I would get customers apologize to me, Eric, we
[9:11] love what you did in that UI, but we're not using
it. Well, why? Well, because we take your data and
[9:17] we take other data and we pull it all together
and we make a much more effective stack and and
[9:22] workflow for how we do things. Fine. We decided
to embrace that paradigm, right? To make sure
[9:28] so the pressure is on to make sure we have great
APIs. What else? To have MCP sitting in front of
[9:33] that to make us clearly cleanly accessible along
with other tools in the in the environment. Jira,
[9:41] right? Uh it might be GitHub as I was mentioning.
It might be my edr, it might be my SIM and so on.
[9:48] So now we can start to orchestrate what's going
on from a security perspective. And [snorts] we're
[9:53] going to interface using coding agents. In a
second, I may show you my desktop just to show
[9:57] you a couple things that I've done. And by the
way, this is not just willy-nilly saying, okay,
[10:03] point claw at this and away you go. One of the key
ingredients here and I'll be interested if anyone
[10:09] has built a skill on their own is that we have
developed and what you're seeing vendors do
[10:15] develop skills that define the best way to not
only access but to use the data the workflows.
[10:24] What kind of additional tools do we recommend
that AI grab hold of? And so from from
[10:31] the moment you start asking questions, the
first thing it does is say, "Let me check.
[10:36] Aha, I've got the cystic skills. Now I know how to
deal with what you're asking me to deal with." It
[10:42] in essence becomes much more security capable for
the tools you have in your shop. In our case, cyig
[10:49] in terms of trying to help you come up
with answers, prioritization solutions,
[10:55] and so skills. Now, anyone built their own
skill for any any reason whatsoever? Yeah,
[10:59] we've got some. A lot of times you'll just
say, "Hey, can I've just asked you all these
[11:03] things. Can we build a skill so that I can reuse
that?" And that's what we're talking about here.
[11:08] Really simple markdown file. When I saw
it was a simple markdown file, I said,
[11:13] "Wow, that's not so hard. I can see what's
actually in it." This is Anyway. Okay.
[11:21] And now we're going to orchestrate our
operations using AI. And I think my last slide,
[11:25] time permitting, we'll talk about how we can
now by necessity move from having to operate
[11:31] everything to just orchestrating things
in our security space. All right. So,
[11:36] what are the advantages? Obviously, automation.
We've been working hard to automate lots of things
[11:42] in in across you name your domain, but especially
in the security space. We want to go faster.
[11:50] And when I don't have to turn
or you don't have to turn to him
[11:53] to say, "Dude, what's I don't understand this.
What's the answer?" Right? Or somebody goes up.
[11:58] So when we don't have to do that, we can be
faster when we can have this, let's call it,
[12:02] an an expert in in our AI engines that we have.
We want to be able to scale. There's so much data
[12:10] flying at us every moment of the day. How do
we even figure out what all that means? And
[12:16] we've seen the impact that AI can have when you've
given it the right data, the right understanding,
[12:24] and helping get from point A to point B in
terms of your your goals. Context matters.
[12:30] I want to be able to customize. And again, one of
the things is I hate doing things this way. Well,
[12:36] we thought that was the right way. That's why
we shipped it in the UI. Well, now you can do it
[12:40] however you want to. building even your own charts
and interfaces and workflows and work it from
[12:47] Slack, whatever you want to do. And there's so
much more that could be put on this list. I know,
[12:52] but I picked a few to help us um just
get some of that on the table. Okay,
[13:00] so what are we doing at Cyig? We took everything.
Anyone familiar with cystic prior to walking in
[13:03] here today? Yeah, some of you. If you were at the
other end of the strip at Black Hat, you probably
[13:08] have more people raise your hand. We're we've got
dueling shows on on e either side of the strip.
[13:14] So we've taken that platform said fine
let's give an interface that will serve
[13:20] when you have your AI coding agent
doing things that require insight
[13:27] in and around security including recommending
to you what to do. Right? A lot of times um
[13:34] we need to remediate a vulnerability. We need to
know how to block an attack path. we need to know
[13:39] all these things and so now we can get help
right and we're doing that via skills MCP the
[13:44] things I the things I mentioned so whether you're
sitting in front of it or whether you're actually
[13:48] creating these agents which I did on my desktop I
said I want an agent that checks every 10 minutes
[13:53] for any new high severity runtime events threats
and uh at first I said send me an email and he
[14:00] goes yeah well that's not a great solution how
about if you have your Mac desktop pop up a
[14:06] an alert I said great do that' So,
five minutes later, I have a working
[14:11] thing that actually looks for the events, pops an
event alert for me, and lets me click and open and
[14:17] get to whatever the insight is. So, very cool way
to have something an agent working on your behalf,
[14:23] doing something for you that you've created to
fit how you like to work. Okay, so what makes Cyic
[14:30] different? I'm going to try and move quickly here
so that I can at least show you a couple things.
[14:35] um you know the data matters and we've always
prided ourselves on having really deep insights
[14:41] into what's happening by looking at kernel level
activity sis calls it's high fidelity so look we
[14:47] all worry about hallucination u whatever it might
be getting wrong conclusions right the best way
[14:54] to get off on the right foot is to give the right
data accurate data to the AI engine to begin with
[15:02] and so without that you might be in for a lot
of guess work, a lot of token spend, a lot of
[15:07] things going on, right? Skills. We mentioned the
ability to personalize and do it the way you want.
[15:12] And the last one is really cool, which is, hey,
I did this today and you asked me questions and
[15:19] it's sort of a dance. You figured out how I like
to run my shop. You figured out when I'm okay
[15:24] and I, you know, hit accept and always accept or
whatever it is, right, that that I'm being asked.
[15:31] Now, as we move forward, do that. Do it this way
because that's what you've learned about me. This
[15:38] is what you've learned about my shop. This is what
you've learned about my Kubernetes and my cloud,
[15:43] right? Whatever it is I'm running,
right? And now that you know that, get my
[15:48] conclusions drawn faster. The way that I like
to see them, the way that I have to report them,
[15:53] the things I'll allow, the things we won't allow.
And there's still plenty of things where we want
[15:58] to hit the pause button and put a human there
to say, "Okay, go ahead and do that." Right?
[16:04] And we acknowledge that full full stop. But
that's all a part of the learning process where
[16:08] things that we're doing in the security space
get much more effective with every interaction.
[16:15] So I mentioned this earlier. We're going
to move from people that have to do this
[16:21] prioritization, watching dashboards, clicking all
the things, swiveling in the chair to talk to to
[16:27] my buddy. Uh, you know, where we have to the only
way we get more done is by adding more headcount
[16:34] and that's a tall tall order today in
many many organizations and enterprises.
[16:40] So instead, we become an orchestrator. We leverage
the tools that we have that are really upon us.
[16:46] It's why we're at this event, right? And yes,
making sure that it understands to ask me before
[16:53] you do anything in my environment and change
anything. I want to be able to also have an audit
[16:58] so that I know what steps have been taken or
what reasoning was done. And now I can scale
[17:05] with my team of agents instead of having to hire
someone every time I need to get a little bit more
[17:11] done. Okay, so we looked at this from a study
standpoint. Just a quick example when we're
[17:17] solving a vulnerability or we're investigating
it. This was inside cydig three different people
[17:23] with three different domain expertise had to
be called upon. Uh they collaborated wasn't
[17:29] that long but 45 it took 45 minutes in all of that
collaboration for them to get to the other side.
[17:36] Now we start adding that up and it means we can
only do so many of these per day per year and it's
[17:41] going to have a cost to it and we just normalize
some um salary information. Now on the on the
[17:48] other side we have someone one person who again
you may not have the knowledge that he or she has
[17:55] but I'm getting help. I'm getting assistance. I'm
asking the questions. 12 minutes later we're done
[18:03] with a cost of about 3 375 in tokens. Not only
does my capacity of doing this every day go up,
[18:09] but it's at a much lower cost. So,
pretty pretty high impact um in terms of
[18:15] when you get this right. Okay. All right. So, I've
got two and a half minutes left, but I just do
[18:20] want to show you real quick what might this look
like. I'm going to give you a couple examples.
[18:24] I was hoping that my little uh alert thing
would would show up, but I don't know that
[18:29] it's going to. So, I've done some creative
things. Like for instance, I just asked for
[18:36] uh Claude to build a chart for me, right? Show me
where in my Kubernetes environment I might have
[18:43] the highest risk and it went out, pulled data from
cyig and I got something that looks like this.
[18:50] Not a lot of expertise in my brain as it relates
to doing these kind of things, but this was one
[18:54] example that I wanted to see visually and
this isn't in the cystic UI as an example.
[19:00] Um, another one that I really like is I've got
things firing in my environment. Show me those.
[19:07] Give me a list. And it's not all about lists and
we know it's about trying to take action, but
[19:11] just to give you examples. And I was pretty amazed
that it's got hover over capabilities. And again,
[19:16] in about seven, eight minutes, I'm understanding
what are the top things firing in my environment.
[19:22] And now it'll ask me usually after this say, okay,
do you want to deal with one or two of these?
[19:26] What do you want to do? Do you want to take
action? Do you want to investigate? And
[19:32] that's where it becomes magical. It's not about
a chatbot. It's about a guide that helps me
[19:39] find out what's what am I most at risk for and
then move me through all the way through things
[19:47] like remediation and response. And I may
not let in my example cloud code do it all,
[19:53] but it's giving me inside and understanding
of what I could do. And when I do want,
[19:58] let's say I've given it the ability to make config
changes in AWS. That looks good. Go for it. Right
[20:06] now, a lot of these things as we build trust in
AI are going to be much more uh effective as we
[20:13] move forward. All right. [snorts] I think we will
pause the demo here because we're out of time, but
[20:17] you can see some of the things here, categories of
investigations and just things that I I was doing
[20:22] throughout uh my time playing with this
as a solution. All right. So 20 seconds
[20:29] the path forward it sh the shift is happening
many of you raised your hands right for being
[20:33] part of security just think about how can this if
you're not doing this today some of you probably
[20:38] are but many of you probably aren't consider how
it's going to help you close gaps or things that
[20:44] you're suffering under today whether that's
vulnerabilities whether that's threats that
[20:48] are moving too fast or or whatever that might look
like and identify a starting point you don't have
[20:52] to do all or nothing and again a lot of our folks
are pointing directly at CVE triage remediation,
[21:00] plugging in Jira, plugging in doing automated
PRs and things like that. That might be a great
[21:05] place to start, right? And you can learn more
more about the approach. I am over time, but uh
[21:11] I do want to thank you for being here and
listening to what we're sort of what we've
[21:16] put together. And this isn't really all
about cyig. This is about how do you take
[21:22] all those tools in your shop and make them work
in a much more orchestrated fashion for you,
[21:28] I like to say, so you can get your
weekends back. All right, that's
