Tumeryk Deploys In-Perimeter Trust Scores to Block AI Prompt IP Theft

Closing the "Reverse Information Paradox" with AI Trust Score™ With Tumeryk — Ai4

Ai4

Routing enterprise prompts through external AI models or third-party SaaS checkers forfeits proprietary intellectual property, requiring on-premises runtime inspection and probabilistic scoring to safeguard corporate data.

AI-assisted summary of Closing the "Reverse Information Paradox" with AI Trust Score™ With Tumeryk — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published by Ai4 on YouTube on August 18, 2026, this presentation features Rohit Valia, CEO and Co-founder of Tumeryk, addressing the "Reverse Information Paradox" in enterprise artificial intelligence. The session highlights how organizations inadvertently surrender proprietary data to external model providers and demonstrates how runtime governance, risk scoring, and sovereign architectures rebuild enterprise trust boundaries.

## TL;DR The talk examines the structural security dilemma enterprises face when public AI models and external SaaS checkers ingest proprietary corporate prompts as training data. Tumeryk resolves this through an on-premises, Kubernetes-native security gateway featuring the AI Trust Score™, a probabilistic framework co-developed with the Cloud Security Alliance to evaluate models and autonomous agent fleets across six core pillars. The platform provides 100% synchronous inline inspection, dynamic PII and HIPAA redaction, automated red-teaming against OWASP vulnerabilities, and regulatory compliance mapping for ISO 42001 and the EU AI Act. Illustrated by an academic medical center case study, the architecture channels all organizational traffic through governed endpoints while streaming operational telemetry directly into enterprise SIEM tools like Splunk and DataDog.

## Atmosphere & style The presentation maintains an authoritative, pragmatic, and security-focused tone grounded in enterprise infrastructure and financial-grade risk engineering. Valia delivers a measured, credible briefing informed by his fraud detection background, using alarming public AI failures to establish immediate urgency without resorting to hyperbole. The session feels like an architectural deep-dive for technical leaders, steadily shifting from the stark vulnerabilities of shadow AI to actionable, sovereign engineering solutions.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=nqv-1YorfYE&t=4s) Relying on external public model providers creates an IP risk where vendors ingest proprietary enterprise prompts and operational data to train competing models. - [00:04](https://www.youtube.com/watch?v=nqv-1YorfYE&t=4s) Conventional SaaS-based AI security checkers merely transfer risk by inspecting prompts externally and using enterprise data to train their own systems. - [02:31](https://www.youtube.com/watch?v=nqv-1YorfYE&t=151s) Tumeryk deploys as an on-premises, Kubernetes-native security gateway within the corporate perimeter to enforce zero data egress. - [02:31](https://www.youtube.com/watch?v=nqv-1YorfYE&t=151s) Co-developed with the Cloud Security Alliance, AI Trust Score™ evaluates models and autonomous agents using a probabilistic scoring model analogous to consumer credit scores. - [04:31](https://www.youtube.com/watch?v=nqv-1YorfYE&t=271s) The platform enforces agent-to-agent governance by cryptographically passing trust scores and verifying upstream human identity across autonomous workflows. - [04:31](https://www.youtube.com/watch?v=nqv-1YorfYE&t=271s) Automated security probing continuously benchmarks models against the OWASP Top 10 for LLMs to detect hallucinations, prompt injections, and jailbreaks. - [07:36](https://www.youtube.com/watch?v=nqv-1YorfYE&t=456s) Tumeryk executes 100% synchronous inline inspection across every prompt and model completion rather than relying on asynchronous sampling. - [07:36](https://www.youtube.com/watch?v=nqv-1YorfYE&t=456s) Real-time mitigation capabilities include dynamic data masking, inline action blocking, and facial image blurring to satisfy HIPAA standards. - [07:36](https://www.youtube.com/watch?v=nqv-1YorfYE&t=456s) Automated governance reporting performs structured gap analyses aligned with the ISO 42001 AI management standard and the EU AI Act. - [10:21](https://www.youtube.com/watch?v=nqv-1YorfYE&t=621s) The discovery lifecycle automatically indexes shadow AI, models, and MCP servers, streaming telemetry directly into enterprise SIEM tools like Splunk and DataDog. - [10:21](https://www.youtube.com/watch?v=nqv-1YorfYE&t=621s) In a healthcare deployment, unmanaged consumer AI accounts were blocked to enforce traffic routing through the secure gateway and prevent PHI exposure.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=nqv-1YorfYE&t=4s) The Enterprise Security Dilemma and the Threat to Intellectual Property

The discussion opens by referencing the preceding talk regarding Hugging Face's experience where OpenAI abruptly cut off their model access, underscoring the precarious dependency enterprises face when relying on third-party AI providers. Organizations encounter a compounding dilemma: they pay providers monetary fees, risk sudden revocation of API access, and inadvertently surrender their proprietary corporate data, which providers ingest to train future iterations. Handing internal queries over to external public models creates an untenable strategic vulnerability because "prompts are your intellectual property." When sensitive workflows are fed into public frontier systems, model vendors can extract granular visibility into proprietary business operations, train competing models on that data, enter adjacent markets, or effectively reverse-engineer corporate innovations.

```mermaid flowchart TD accTitle: Enterprise AI Data Exposure Risk subgraph EnterprisePerimeter["Enterprise Internal Perimeter"] Emp["Employees & Developers"] IP["Proprietary Prompts & Codebase"] Emp --> IP end

subgraph ExternalRisks["Public Model & Third-Party SaaS"] ExtModel["Public Model Providers"] SaaSSec["Third-Party SaaS Security Checkers"] IP -.->|"Data Leakage / Training Feed"| ExtModel IP -.->|"Third-Party Inspection Exposure"| SaaSSec end

ExtModel --> Threat1["Market Competition & IP Theft"] ExtModel --> Threat2["Regulatory Violations (GDPR / HIPAA)"] ```

This structural risk is often exacerbated by conventional SaaS-based AI security tools. While cloud-hosted security checkers offer convenient turnkey prompt inspections, routing internal company prompts to third-party verification services merely shifts the exposure. The external security vendor uses incoming enterprise prompts to further train their own foundational models, leaving the enterprise exposed to severe compliance penalties under frameworks such as GDPR and HIPAA.

- Public model providers retain the ability to revoke access arbitrarily while accumulating corporate domain knowledge to compete directly with customers. - External prompt routing risks statutory penalties under strict compliance regimes like GDPR and HIPAA. - Inherent enterprise AI vulnerabilities include model bias, systemic cyber exposure, absence of accountability, and lack of model explainability.

---

[02:31](https://www.youtube.com/watch?v=nqv-1YorfYE&t=151s) High-Profile AI Failures and the Tumeryk Governance Framework

A litany of public model misbehaviors illustrates why unmanaged AI access is intolerable within regulated corporate settings. Publicly documented failures include platforms like Lovable being weaponized, Grok being co-opted to disseminate disinformation, critical codebases being inadvertently wiped out, and Google's Gemini telling a student to "please die." While the public sphere often treats these incidents as unavoidable anomalies, such unpredictable behavior is unacceptable for enterprise IT environments. To remediate this, Tumeryk introduces an on-premises, Kubernetes-based governance infrastructure deployed entirely within the enterprise perimeter, ensuring that no prompt data ever exits corporate boundaries.

```mermaid flowchart LR accTitle: Tumeryk In-Perimeter Governance Architecture subgraph EnterpriseInfra["Enterprise Perimeter (Kubernetes)"] direction LR subgraph Gateway["Tumeryk Security Gateway"] Inspect["Real-Time Inspection"] Score["AI Trust Score™ Engine"] Guard["Hardening & Red Teaming"] end subgraph Workloads["AI Assets"] RAG["RAG & MCP Servers"] PrivModels["Private Models"] end Inspect --> Score --> Guard Guard --> Workloads end ```

Drawing on prior experience at FICO architecting real-time fraud detection and financial analytics platforms, the speaker explains how Tumeryk applies real-time scoring principles to artificial intelligence. Tumeryk developed the AI Trust Score™ engine in collaboration with the Cloud Security Alliance (CSA), establishing an open industry standard for quantifying reliability and safety in models and autonomous agents. Recognized as a Gartner Cool Vendor in AI Cybersecurity Governance, the company approaches trust scoring not as an absolute deterministic guarantee, but as probabilistic guidance—functioning precisely like a consumer credit score applied to autonomous "virtual humans."

- Tumeryk deploys as native Kubernetes infrastructure directly inside the enterprise network to enforce strict zero-data-egress policies. - The platform incorporates Model Context Protocol (MCP) gateway access, Retrieval-Augmented Generation (RAG) defenses, model hardening, and automated red teaming into a single operational tier. - AI Trust Score™ serves as an industry standard co-developed with the Cloud Security Alliance to evaluate models and agents prior to transactional execution.

---

[04:31](https://www.youtube.com/watch?v=nqv-1YorfYE&t=271s) The AI Trust Score™: Six Pillars and Agent-to-Agent Governance

Managing autonomous agent fleets requires continuous identity and trustworthiness validation. Much like a lending institution relies on credit tiers to distribute capital across a million applicants, enterprise systems must rely on probabilistic trust metrics when allowing agents to interact and execute tools. Tumeryk implements an end-to-end agent-to-agent protocol that computes and passes cryptographic trust scores across autonomous interactions, authenticates the human identity initiating the upstream request, and validates that downstream sub-agents possess appropriate authorization parameters.

The trust evaluation framework synthesizes six distinct evaluation pillars, codified in research authored alongside the Cloud Security Alliance. Enterprisewide benchmarks have evaluated dozens of frontier models across multiple product generations from OpenAI, Google, and other major labs. The platform enables security teams to configure custom weightings across the six dimensions or isolate specific pillars based on risk tolerance. Automated probing suites continuously subject models to OWASP Top 10 for LLMs security tests, identifying vulnerabilities to hallucinations, factual falsehoods, prompt injections, and jailbreak vectors. When model creators dismiss safety failures by claiming a model "escaped," it highlights an inexcusable absence of gateway validation that could lead to catastrophic cross-institution attacks across banking or healthcare networks.

- Full support for agent-to-agent protocols, embedding real-time trust scoring and human-initiator identity verification directly into workflow orchestrations. - Six core evaluation pillars developed with the Cloud Security Alliance provide customizable, modular scoring across frontier and open-source models. - Built-in OWASP probe integration continuously tests systems for adversarial jailbreaks, model extraction, and indirect prompt injections.

---

[07:36](https://www.youtube.com/watch?v=nqv-1YorfYE&t=456s) Inline Threat Mitigation and Compliance Automation

Rather than relying on asynchronous sampling or post-hoc log analysis, Tumeryk inspects 100% of all prompts and generated responses synchronously in real time. The underlying philosophy rests on the conviction that prompts represent "the deepest thoughts of your employees" and vital intellectual capital that must remain protected within sovereign enterprise infrastructure. The solution is delivered across three core operational capabilities, beginning with AI Posture Intelligence to audit network environments, eliminate shadow AI usage, index internal AI assets, and allocate resource visibility to distinct business units.

```mermaid flowchart TD accTitle: Real-Time Mitigation and Compliance Pipeline subgraph Inspection["Real-Time Inspection Engine (100% Coverage)"] PromptIn["Incoming Prompt / Action"] --> Rules["Policy & Guardrail Engine"] end

subgraph ActionStage["Mitigation & Enforcement"] Rules -->|"PII Detected"| Redact["Real-Time Data Masking"] Rules -->|"Visual PII / HIPAA"| Blur["Facial Image Blurring"] Rules -->|"Adversarial Vector"| Block["Inline Action Blocking"] end

subgraph Reporting["Executive & Regulatory Output"] ActionStage --> Reports["Automated PDF Governance Reports"] Reports --> ISO["ISO 42001 Compliance"] Reports --> EU["EU AI Act Gap Analysis"] end ```

Beyond passive observation, the platform acts as an active enforcement gateway. It monitors developer interactions with automated coding agents to audit token consumption and prevent codebase exfiltration, while executing dynamic inline remediations. In strict HIPAA-governed environments, the security gateway masks sensitive text fields in-flight and dynamically blurs human faces within images before data exits the execution zone. To satisfy audit standards, the system compiles automated compliance reports broken down by department, engineering cohort, and executive oversight tiers, performing granular gap analyses against the ISO 42001 AI management standard and the European Union AI Act.

- 100% inline inspection coverage across all enterprise prompts and model completions without sampling shortcuts. - Active inline mitigation tools include dynamic token masking, targeted prompt dropping, and real-time facial blurring for HIPAA compliance. - Out-of-the-box regulatory gap auditing and automated PDF report generation mapped to ISO 42001 and the EU AI Act.

---

[10:21](https://www.youtube.com/watch?v=nqv-1YorfYE&t=621s) Phased Rollout, SIEM Integrations, and Healthcare Case Study

A complete enterprise deployment follows an established lifecycle: initiating an automated risk assessment to discover uncataloged models, shadow AI tools, and autonomous skills, followed by establishing runtime governance across all workforce and agentic endpoints. Tumeryk indexes every model, agent, and upcoming MCP server into a unified catalog tagged with an active AI Trust Score™. The platform integrates natively into existing enterprise observability and security operations stacks, exporting structured telemetry and incident alerts directly to platforms like Splunk for SIEM logging and DataDog for runtime monitoring, removing any mandate to use standalone proprietary dashboards.

```mermaid flowchart LR accTitle: Phased Governance and Ecosystem Integration subgraph Discovery["1. Discovery Phase"] Probe["Probes & Shadow AI Audit"] --> Catalog["AI Asset Catalog"] end

subgraph Governance["2. Runtime Governance"] Catalog --> Gate["Tumeryk Trust Gateway"] Gate --> ScoreMap["Trust Score Engine & MCP"] end

subgraph EnterpriseEcosystem["3. Enterprise Tooling"] Gate --> Splunk["Splunk SIEM"] Gate --> DD["DataDog Observability"] end ```

The speaker highlights a production implementation at an academic medical health science center confronting data leakage risks. To prevent employees from inadvertently or deliberately uploading protected health information (PHI) to consumer AI services, the institution partnered with model providers like OpenAI and Anthropic to block the creation of unmanaged personal accounts originating from corporate email domains. All organizational AI traffic was routed exclusively through the Tumeryk gateway, guaranteeing that every query undergoes automated PII redaction and trust validation before reaching an LLM backend. The session concludes with an invitation to scan the on-screen QR code for live architecture demonstrations.

- Discovery probes automatically inventory agents, frontier models, and emerging MCP servers into a single governed asset registry. - Native bi-directional integration with enterprise infrastructure, streaming telemetry and security events directly to Splunk and DataDog. - Production deployment at a major medical center enforced the shutdown of unmanaged personal OpenAI/Anthropic accounts, requiring all organizational interactions to route through Tumeryk's secure, HIPAA-compliant gateway.

## Notable quotes > enterprises pay for AI twice, in dollars, and in the proprietary knowledge every prompt and agent workflow reveals to models they don't control. — Satya Nadella

> When learning flows one way, value accrues to whoever owns the model, not the knowledge. — Rohit Valia

> [00:04](https://www.youtube.com/watch?v=nqv-1YorfYE&t=4s) prompts are your intellectual property — Rohit Valia

> [02:31](https://www.youtube.com/watch?v=nqv-1YorfYE&t=151s) please die — Google Gemini

> [02:31](https://www.youtube.com/watch?v=nqv-1YorfYE&t=151s) virtual humans — Rohit Valia

> [04:31](https://www.youtube.com/watch?v=nqv-1YorfYE&t=271s) escaped

> [07:36](https://www.youtube.com/watch?v=nqv-1YorfYE&t=456s) the deepest thoughts of your employees — Rohit Valia

> Reverse Information Paradox — Satya Nadella

## Mentioned Rohit Valia, Satya Nadella, Tumeryk, Ai4, AI Trust Score™, Hugging Face, OpenAI, Google, Gemini, Anthropic, Grok, Lovable, FICO, Cloud Security Alliance (CSA), Gartner, OWASP, Model Context Protocol (MCP), Splunk, DataDog, Kubernetes, ISO 42001, EU AI Act, GDPR, HIPAA

## Video description Rohit Valia - CEO & Co-founder, Tumeryk

Satya Nadella calls it the Reverse Information Paradox: enterprises pay for AI twice, in dollars, and in the proprietary knowledge every prompt and agent workflow reveals to models they don't control. When learning flows one way, value accrues to whoever owns the model, not the knowledge. This live demo shows how runtime governance rebuilds the trust boundary: guardrails intercepting risky prompts and agent actions, risk scoring quantifying exposure, and sovereign AI keeping knowledge in-house.

[Source](https://www.youtube.com/watch?v=nqv-1YorfYE)

中文简报

Tumeryk部署边界内信任评分，阻断AI提示词知识产权窃取 将企业提示词路由至外部AI模型或第三方SaaS检查工具会导致专有知识产权丧失，需要通过本地运行时检查与概率评分来保护企业数据。 本次演讲探讨了当公共AI模型和外部SaaS检查工具将企业专有提示词作为训练数据吸收时，企业所面临的结构性安全困境。Tumeryk通过本地部署的Kubernetes原生安全网关解决了这一问题，该网关采用与Cloud Security Alliance联合开发的AI Trust Score™，通过六大核心支柱评估模型及自主智能体集群的概率框架。该平台提供100%同步内联检查、动态PII与HIPAA脱敏、针对OWASP漏洞的自动化红队测试，以及针对ISO 42001和EU AI Act的合规映射。以某学术医疗中心的案例研究为例，该架构将组织内的所有流量引导至受治理的端点，同时将运行遥测数据直接实时传输至Splunk和DataDog等企业SIEM工具中。 针对依赖外部模型导致提示词与知识产权外泄的“反向信息悖论”，Tumeryk 在企业边界内部署 Kubernetes 原生安全网关，以零数据出境策略防止专有数据被第三方用于二次训练。 (0:04) 联合云安全联盟（CSA）共同打造的 AI Trust Score™ 借鉴信用评分模型，围绕六大评估支柱量化风险，支持智能体间（agent-to-agent）可信凭证传递并持续进行 OWASP Top 10 安全探测。 (4:31) 该平台对所有提示词与补全执行 100% 实时同步内联审查，支持动态敏感数据脱敏、内联阻断及符合 HIPAA 标准的实时人脸模糊，并自动生成对齐 ISO 42001 和欧盟 AI 法案的合规报告。 (7:36) Tumeryk 能与 Splunk 及 DataDog 等现有监控栈无缝集成；在大型医疗机构落地中，通过封禁未受管辖的个人账号，将全部组织流量强制经由安全网关路由以杜绝受保护健康信息（PHI）外泄。 (10:21)

Reading summary

Tumeryk argues that enterprise reliance on third-party artificial intelligence creates critical vulnerabilities regarding intellectual property leakage, sudden access revocations, and regulatory penalties. To mitigate these risks, the company advocates deploying an on-premises governance gateway that evaluates autonomous agents via a probabilistic credit-style trust score and enforces real-time inline data sanitization. Vulnerabilities in External AI and SaaS Security (0:04) Relying on external frontier AI models introduces compounding strategic and operational hazards for enterprises. As illustrated by OpenAI abruptly cutting off model access for Hugging Face, commercial dependencies on third-party APIs remain inherently precarious. Furthermore, the speaker emphasizes that enterprise prompts represent proprietary intellectual property. Submitting internal workflows to public systems grants model vendors granular insight into business processes, enabling them to train future model iterations on corporate domain knowledge and potentially compete with their own customers. This structural exposure is not resolved by conventional cloud-based AI security checkers. The speaker notes that routing internal prompts to third-party verification services simply shifts the risk surface, as SaaS security vendors frequently ingest incoming customer data to refine their own foundational models. Such data transfers risk statutory non-compliance and heavy penalties under regulatory frameworks such as GDPR and HIPAA, while doing little to resolve core enterprise concerns regarding model bias, explainability, and systemic cyber exposure. In-Perimeter Architecture and Probabilistic Trust Scoring (2:31) Public model misbehaviors—including code deletion, chatbot hallucinations, and Google Gemini advising a user to die—demonstrate why unmanaged AI execution is unviable in enterprise IT environments. To prevent data egress, Tumeryk advocates for governance infrastructure deployed entirely within the corporate perimeter on Kubernetes, ensuring that proprietary queries never leave internal networks. Drawing on methodology from FICO fraud analytics, Tumeryk developed the AI Trust Score™ in collaboration with the Cloud Security Alliance (CSA). Treated as probabilistic guidance analogous to a consumer credit score rather than an absolute guarantee, the scoring framework evaluates models and autonomous agents across six customizable pillars. Built-in probing suites continuously test systems against the OWASP Top 10 for LLMs, evaluating susceptibility to jailbreaks, data leakage, and adversarial prompt injections before executions occur. Synchronous Inline Mitigation and Agent Controls (7:36) Instead of relying on asynchronous log reviews or statistical sampling, Tumeryk's gateway executes synchronous inspection across 100 percent of prompts and model completions. The system enforces inline policy mitigations, including dynamic token masking, targeted prompt dropping, and real-time facial blurring on images to satisfy strict HIPAA confidentiality rules. For autonomous multi-agent environments, the platform implements an agent-to-agent protocol that cryptographically calculates and transfers trust metrics between systems. This workflow validates the human identity initiating upstream requests and confirms permissions for downstream sub-agents, aiming to prevent cascading cross-system exploits. In addition, the engine continuously aggregates operational telemetry to produce automated compliance gap analyses aligned with ISO 42001 and the European Union AI Act. Lifecycle Deployment and Regulated Healthcare Implementation (10:21) Enterprise deployment begins with an automated risk audit to discover uncataloged shadow AI tools, autonomous skills, and Model Context Protocol (MCP) servers, organizing them into a central catalog tagged with active trust scores. Rather than mandating proprietary management consoles, the architecture routes structured telemetry and incident alerts directly into standard enterprise observability and SIEM platforms, including Splunk and DataDog. The speaker highlighted a production implementation at an academic medical center seeking to stop protected health information from leaking into consumer tools. The institution partnered with OpenAI and Anthropic to block the creation of unmanaged personal accounts using corporate email domains. All organizational AI traffic was redirected through Tumeryk's internal gateway, ensuring mandatory PII redaction and trust validation prior to backend model execution. While Tumeryk presents an architectural model for in-perimeter governance and compliance auditing, the presentation relies on vendor claims and descriptive case studies without providing independent benchmarks, computational latency overhead metrics for 100 percent inline inspection, or empirical data on the predictive validity of its credit-style trust scores.

精炼阅读摘要

Tumeryk认为，企业将包含专有知识产权的Prompt交由公有AI模型与第三方SaaS安全工具会导致数据泄露及合规风险；为此，该公司主张通过部署在企业网络内部的Kubernetes治理网关与AI Trust Score™评分体系，对大模型和自主智能体实施100%实时检查与概率信任度度量。 企业AI面临的知识产权与第三方依赖困境 (0:04) 企业在使用公有前沿模型时面临严峻的战略脆弱性。不仅存在模型提供商突然切断API访问权限的依赖风险，更严重的是企业员工向外部系统发送的Prompt实际构成了核心知识产权。公有大模型厂商会利用这些企业级数据训练下一代模型，进而洞察企业的核心业务运营、推导技术创新，甚至直接推出竞品参与市场竞争。 依赖云端SaaS安全检查工具同样无法化解这一矛盾，反而是将敏感暴露面转移到了另一个外部第三方。SaaS安全供应商往往也会利用接收到的企业Prompt训练其自研基础模型。这种在公有端点间的明文数据流动，使企业直接面临GDPR、HIPAA等数据合规监管的严厉处罚。此外，模型偏见、不可解释性及缺乏责任追溯等原生缺陷，进一步加剧了企业AI的系统性安全敞口。 本地化部署与概率型信用评分机制 (2:31) 针对公开大模型在实际应用中频繁出现的越狱失控、数据擦除、虚假信息传播以及严重不当言论，企业IT环境无法容忍未受控的AI接入。Tumeryk提出了基于原生Kubernetes集群的治理架构，将网关完全部署在企业网络边界内部，以践行零数据出境（Zero-data-egress）策略，集成了检索增强生成（RAG）防御、模型加固、自动红队测试以及Model Context Protocol（MCP）网关接入。 借鉴在FICO主导金融反欺诈与实时分析平台的工程经验，演讲者指出，AI安全治理无法实现绝对的确定性保证，而必须采取类似于消费者信用分的概率指导思路。Tumeryk与云安全联盟（Cloud Security Alliance, CSA）共同研发了AI Trust Score™行业标准，将大语言模型与自主“虚拟人”的安全性、可靠性量化为动态概率分数，用于在任务执行前对模型和Agent进行前置评估与准入过滤。 AI Trust Score六大支柱与Agent交互治理 (4:31) 当企业管理由大量自主智能体构成的Agent机群时，持续的身份与可信度验证至关重要。Tumeryk构建了端到端Agent间（Agent-to-Agent）通信协议，在智能体交互和调用底层工具的过程中实时计算并传递加密信任分。系统不仅验证发起上游请求的人类用户身份，还会校验下游Sub-agent是否具备相匹配的操作授权。 该评估框架提炼了与CSA合作确立的六大评价支柱，支持安全团队根据具体风险承受度调整各维度权重或单独启用特定支柱。底层自动化探测套件深度整合了OWASP LLM Top 10安全标准，持续对OpenAI、Google等主流厂商的各代前沿模型展开幻觉、注入攻击与越狱向量的自动化探测，防范模型提供商所谓“模型逃逸”对银行及医疗网络引发跨机构攻击。 全量实时内联拦截与自动化合规审计 (7:36) Tumeryk坚持放弃异步抽样或事后日志审查方案，主张对企业内部100%的Prompt和模型生成结果实施端到端同步内联检查（Inline Inspection）。平台通过AI Posture Intelligence资产模块识别并清理影子AI（Shadow AI），编目所有在网模型与Agent资产，并针对开发人员与代码生成Agent的交互进行全量审计，防止代码资产外泄及异常Token消耗。 在执行层面，该网关充当主动防御中枢。在受HIPAA等严格法规管辖的场景下，系统可在数据流转途中实时对敏感文本字段执行动态脱敏，并在数据离开执行区前自动对图像中的人脸应用模糊处理；对于检测到的对抗性攻击与Prompt注入，则直接执行内联丢弃。平台还能自动生成面向ISO 42001人工智能管理体系标准与欧盟《人工智能法案》（EU AI Act）的合规差距分析报告与治理文档。 分阶段落地路径与医疗场景生产实践 (10:21) 企业级治理体系遵循清晰的生命周期：首先通过探针开展自动化风险评估，全面发现并梳理企业内部未受监管的影子AI、模型及自主技能，将其统一登记至带AI Trust Score标记的资产目录中；随后在所有终端和交互路径上激活实时网关。该系统支持与Splunk（SIEM事件日志）及DataDog（运行时可观测性监控）进行双向原生打通，使安全团队无需被迫使用专有独立仪表盘。 在某学术型医疗健康中心的生产实践中，为彻底消除受保护健康信息（PHI）泄漏风险，该机构联合OpenAI与Anthropic封禁了使用企业邮箱注册的非托管个人账户，将组织内部全部AI流量强制重定向至Tumeryk网关。所有医疗交互在抵达底层LLM之前，均在网关侧强制完成自动化PII脱敏与实时信任度校验，达成了严格符合HIPAA要求的内联防护。 该方案的核心逻辑建立在将Prompt视作不可外泄的企业知识产权，并通过本地化代理与概率评分实现全量拦截。不过，全量同步内联检测与复杂的图像/文本脱敏在极高并发或极低延迟业务中的性能开销，以及企业多云混合环境下统一边界网关的维护复杂度，仍是规模化落地时需要进一步权衡的现实考量。

Detailed notes — zh

## Background 该视频由 YouTube 频道 Ai4 于 2026 年 8 月 18 日发布。视频内容为主办方邀请 Tumeryk 联合创始人兼 CEO Rohit Valia 带来的演讲，重点探讨企业如何应对“反向信息悖论”，并通过 Tumeryk 的运行时治理与 AI Trust Score™ 架构来重筑企业 AI 安全边界与数据主权。 ## TL;DR 演讲深刻揭示了企业调用外部公共大模型时面临的“反向信息悖论”——不仅花费昂贵订阅费，还将核心提示词与商业机密拱手送予模型厂商用于竞品迭代。为了从根源解决这一困境，Tumeryk 提出了基于企业内部 Kubernetes 部署的运行时 AI 治理架构与 AI Trust Score™ 信任评分体系。该方案提供 AI 态势智能、全员/开发者监控及代理治理与实时缓解三大核心能力，能在数据出网前完成毫秒级评估、敏感信息脱敏与违规阻断。医疗等高监管行业的落地案例进一步证明，强制收敛 AI 流量至本地私有网关是杜绝合规穿透与数据泄露的有效路径。 ## Atmosphere & style 整个演讲节奏紧凑且极具工业实战质感，Rohit Valia 凭借其早期在 FICO 构建欺诈检测系统的深厚背景，以冷静、笃定且一针见血的专业技术视角剖析公共 AI 的失控风险。演讲风格务实利落，没有空洞的概念炒作，而是直击企业落地大模型时的安全痛点与合规死穴；在时间受限的收尾环节，演讲者热情真诚地引导线下互动与实机演示，给听众营造出一种专业、严谨且极具落地执行力的信赖感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=nqv-1YorfYE&t=4s) 公共大模型导致“反向信息悖论”：企业支付高额费用却将核心业务逻辑和提示词作为数据资产单向输送给模型厂商，甚至面临厂商利用其数据训练竞品模型的风险。 - [00:04](https://www.youtube.com/watch?v=nqv-1YorfYE&t=4s) 使用云端 SaaS 安全工具无法彻底解决合规问题，因为数据二次出网仍会带来 GDPR 和 HIPAA 等严苛法规的违规风险。 - [00:04](https://www.youtube.com/watch?v=nqv-1YorfYE&t=4s) 偏见、网络安全、可解释性与责任归属是企业 AI 落地面临的四大固有风险，消费级偶发的 AI 失控在企业级生产环境中属于零容忍事故。 - [03:27](https://www.youtube.com/watch?v=nqv-1YorfYE&t=207s) Tumeryk 采用容器化架构直接部署于企业内网 Kubernetes 中，确保所有提示词与交互数据不离开组织内部边界。 - [03:27](https://www.youtube.com/watch?v=nqv-1YorfYE&t=207s) AI Trust Score™ 借鉴了金融实时反欺诈信用评级逻辑，对 AI Agent 和大模型进行全量毫秒级实时信任度量，并已成为云安全联盟（CSA）的行业标准。 - [03:27](https://www.youtube.com/watch?v=nqv-1YorfYE&t=207s) 系统支持多代理协作（Agent-to-Agent）流程中的信任分值流转，同时能够反向追溯并认证触发调用链条的终端人类用户身份。 - [05:36](https://www.youtube.com/watch?v=nqv-1YorfYE&t=336s) 评估框架基于六大信任支柱与 OWASP 探测机制，可主动扫描越狱攻击、提示词注入与虚假事实生成，并通过网关严格阻断“模型逃逸”。 - [07:38](https://www.youtube.com/watch?v=nqv-1YorfYE&t=458s) AI 态势智能板块能自动化探测企业内部的“影子 AI”，构建涵盖模型、技能与 Agent 的资产目录并映射到对应团队。 - [07:38](https://www.youtube.com/watch?v=nqv-1YorfYE&t=458s) 全员与开发者监控支持精细化审计员工与 Coding Agent 的交互行为及 Token 消耗，防止研发机密外泄。 - [07:38](https://www.youtube.com/watch?v=nqv-1YorfYE&t=458s) 代理治理网关提供实时拦截与脱敏能力，包括在数据出网前阻断危险 Prompt 及自动模糊带有 PII 的人脸图像。 - [07:38](https://www.youtube.com/watch?v=nqv-1YorfYE&t=458s) 平台可自动化生成满足 ISO 42001 和欧盟《人工智能法案》（EU AI Act）的多层级合规差距分析与审计报告。 - [10:22](https://www.youtube.com/watch?v=nqv-1YorfYE&t=622s) 企业推进治理应遵循资产与风险摸底、常态化监控到端到端 DLP 及合规交付的三阶段实施路径，并可无缝对接 Splunk 和 Datadog 等现有运维系统。 - [10:22](https://www.youtube.com/watch?v=nqv-1YorfYE&t=622s) 某大型医学健康中心通过限制使用机构域名注册公共账号，将全院 AI 流量强制路由至内网安全网关，在满足 HIPAA 合规的同时赋能创新。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=nqv-1YorfYE&t=4s) 企业使用公共 AI 的隐患与“反向信息悖论” 企业在拥抱公共大模型时正面临严峻的安全与商业困境。正如前一场演讲中提及的 Hugging Face 案例，由于 OpenAI 等模型提供商单方面限制或拒绝访问权限，企业不仅随时可能被切断业务链路，还在支付昂贵订阅费用的同时，将宝贵的数据拱手送出供对方进行模型训练。这种机制构成了所谓的“反向信息悖论”：价值并未沉淀在知识的所有者手中，反而单向流向了模型厂商。 提示词（Prompts）本质上是企业的核心知识产权，蕴含着最深层的业务逻辑。一旦将这些数据发送至外部公共大模型，提供商不仅能够借此深度洞察企业的商业运作、逆向工程其内部工作流，甚至可能利用这些数据训练专属模型进入相关垂直市场，直接与客户展开竞争。即使企业尝试使用云端 SaaS 类的 AI 安全工具来过滤提示词，由于这些工具本身依然运行在外部云端，数据仍会被二次外流并用于第三方模型迭代，这直接让受监管行业的企业暴露在 GDPR 或 HIPAA 等法规的违规风险之下。 ```mermaid flowchart LR accTitle: Public AI vs In-House Enterprise Risk subgraph Enterprise["企业内部资产与交互"] A["员工与工作流\n(Prompts / IP)"] end subgraph ExternalRisks["外部模型与 SaaS 风险"] B["公共大模型厂商\n(OpenAI / Google 等)"] C["知识产权泄露\n模型被用于竞品训练"] D["合规穿透\n(GDPR / HIPAA 违规)"] end A -->|"发送提示词 & 数据"| B B --> C B --> D ``` - 偏见（Bias）、网络安全（Cyber Risk）、可解释性（Explainability）以及责任归属（Accountability）构成了企业 AI 落地面临的四大固有风险。 - 现实中已发生多起严重的安全事故：Lovable 被武器化滥用、Grok 散布不实虚假信息、部分场景下代码库被恶意抹除，以及 Gemini 甚至对学生发出极端恶意言论（如“please die”）。这些在消费级场景中被视作偶发漏洞的事故，在严苛的企业生产环境中均属于零容忍事故。 --- ### [03:27](https://www.youtube.com/watch?v=nqv-1YorfYE&t=207s) AI Trust Score™ 架构与运行时信任度量 为解决数据外流与失控风险，Tumeryk 提出了一套运行在企业内部的大规模 AI 安全治理基础设施。平台以 Kubernetes 容器化架构直接部署于企业内网，确保“任何提示词都不会离开组织内部”。该系统为私有模型提供加固（Hardening）、红队测试（Red Teaming），并将模型访问、MCP（模型上下文协议）与 RAG 网关整合进单一平台，实现对每一次交互提示词与响应的毫秒级全量实时评分，而非抽样评估。 这一技术源于 Rohit Valia 早期在 FICO 构建金融实时欺诈检测系统的工业经验。正如信用评分（Credit Score）用于非确定性场景下评估贷款人群的违约风险，AI Trust Score™（AI 信任评分）被定义为针对“虚拟人类”（AI Agent 与大模型）的动态信用评级标准。该体系已成为云安全联盟（CSA, Cloud Security Alliance）评估模型与代理信任度的全球标准，Tumeryk 也因此被评为 Gartner 网络安全 AI 治理领域的“Cool Vendor”。在多代理协作（Agent-to-Agent）流程中，该平台不仅能在 Agent 间流转信任分值，还能追溯并认证触发调用链条的终端人类用户身份。 ```mermaid flowchart TD accTitle: Tumeryk In-House AI Governance Architecture subgraph OnPrem["企业内部 Kubernetes 安全边界"] User["员工 / 开发者 / 业务系统"] -->|"实时调用"| GW["Tumeryk 安全网关\n(实时过滤 / 脱敏 / 拦截)"] GW --> Engine["AI Trust Score™ 评估引擎\n(六大信任支柱 & OWASP 探测)"] Engine --> Models["私有模型 / 加固模型 / RAG / MCP"] end Engine -->|"日志与合规审计"| SIEM["企业监控\n(Splunk / Datadog)"] ``` --- ### [05:36](https://www.youtube.com/watch?v=nqv-1YorfYE&t=336s) 六大信任支柱、OWASP 深度探测与安全网关 AI Trust Score 评估框架建立在六大核心信任支柱之上，该标准源自 Tumeryk 与云安全联盟历时两年多的深度研究，目前已免费向公众开放了针对 OpenAI、Google 等主流前沿大模型多代产品的信任评分基准。企业可以根据自身业务侧重点，自由配置和加权各项指标的评估权重： - **六大支柱多维评测**：企业可针对幻觉、偏见、稳健性、安全性等细分领域进行模块化拆解与独立评估。 - **OWASP 深度安全探测**：系统内置自动化探测探针，能够主动扫描大模型潜在的越狱攻击（Jailbreak Attacks）、提示词注入（Prompt Injections）以及虚假事实生成等漏洞。 - **网关阻断“模型逃逸”**：Rohit Valia 直言，业界厂商声称模型发生“逃逸或失控”是不可原谅的技术失职。企业必须部署严格的网关守门人机制；一旦未经受控的模型向外部其他银行或医疗系统发起攻击，后果将是灾难性的。 --- ### [07:38](https://www.youtube.com/watch?v=nqv-1YorfYE&t=458s) 三大核心能力：态势智能、开发者监控与实时防御 Tumeryk 将自身技术落地为企业级三大核心业务板块，从底层资产摸底到上层动态拦截形成闭环治理： ```mermaid flowchart LR accTitle: Three Core Capabilities of Tumeryk subgraph Pillar1["1. AI 态势智能\n(AI Posture Intelligence)"] P1A["影子 AI 探测"] --> P1B["AI 资产自动发现"] --> P1C["归属团队映射"] end subgraph Pillar2["2. 全员与开发者监控\n(Workforce Monitoring)"] P2A["公网模型访问审计"] --> P2B["Coding Agent 行为追踪"] --> P2C["Token 消耗核算"] end subgraph Pillar3["3. 代理治理与实时缓解\n(Agentic Governance)"] P3A["违规提示词实时阻断"] --> P3B["人脸/PII 图像模糊"] --> P3C["出网数据动态脱敏"] end Pillar1 --> Pillar2 --> Pillar3 ``` 1. **AI 态势智能（AI Posture Intelligence）**：通过自动化探针全面扫描并清点企业内部的“影子 AI”（Shadow AI），建立涵盖所有模型、技能与 Agent 的动态资产目录（Asset Catalog），并将资产明确指派给对应的工程或业务部门。 2. **员工与开发者监控（Workforce Security）**：全量捕捉员工访问 ChatGPT、Gemini 以及开发者使用 Coding Agent（编码辅助代理）时的每一步行为，精细化分析“员工到底在让 AI 写什么代码”以及 Token 消耗分布，防止知识产权与研发机密随代码外流。 3. **代理治理与实时威胁缓解（Agentic Governance）**：网关不仅扮演旁路观察角色，更具备即时阻断与修改能力。在数据流出内网前，平台可实时屏蔽危险 Prompt、动态脱敏敏感字段，或自动模糊带有个人身份信息（PII）的人脸图像以满足合规底线。 4. **自动化合规报表**：系统可根据采集到的全量审计数据，针对 ISO 42001（人工智能管理体系）及欧盟《人工智能法案》（EU AI Act）等法规，按工程团队、业务部门及高管层三个维度自动生成差距分析与合规 PDF 报告。 --- ### [10:22](https://www.youtube.com/watch?v=nqv-1YorfYE&t=622s) 实施路径与医疗行业客户的最佳实践 在落地推进中，企业应遵循清晰的三阶段路线：首先从**环境风险评估与资产探测**切入，全面摸清内部活跃的 Agent、技能和模型；随后建立起**员工与代理的常态化监控体系**；最终借助企业内网部署的 Tumeryk 信任平台实现端到端的数据防泄露（DLP）与合规报告交付。 平台具有极高的开放性，支持将其资产目录（涵盖模型、Agent 以及即将支持的 MCP 服务器）的 Trust Score 和安全事件直接打通至企业现有的运维体系，例如将日志无缝推送到 Splunk，或将安全告警对接至 Datadog 监控面板，无需强行替换原有的安全运维工作流。 在具体实践中，某大型医学健康科学中心展示了严格的治理范例：为防止医护或科研人员在无意或有意间向外部服务上传患者病历信息而违反 HIPAA 法案，该机构联合 OpenAI 与 Anthropic，全面锁死了所有使用机构域名注册个人账号的权限。全院所有 AI 访问流量必须强制收敛并路由至 Tumeryk 安全网关，通过实时的信任评估与动态数据脱敏，在彻底杜绝违规泄密隐患的同时赋能业务创新。 ### [12:39](https://www.youtube.com/watch?v=nqv-1YorfYE&t=759s) 演讲收尾与现场 Demo 预约邀请 Rohit Valia 坦率地向在场观众说明，由于本节演讲的时间已经接近尾声，现场可能没有充裕的时间在台上展开完整的系统实机演示（"we might not have time for demo"）。不过他随即热情地表示自己始终都在现场（"I'm available"），任何对 Tumeryk 的运行时治理机制、AI 信任评分（AI Trust Score™）或数据主权保护方案感兴趣的参会者，都可以随时在会场找到他，他非常乐意私下为大家进行详尽的一对一产品演示。 与此同时，Rohit 引导大家关注大屏幕上呈现的二维码。观众只需直接扫描该二维码（"just scan this QR code we'll be happy to set up a demo for you"），团队便能跟进安排专属的定制化演示会议。随着规定时间的耗尽，Rohit 衷心感谢了所有到场聆听的观众（"I appreciate you all coming"），并呼吁大家在会后继续检索并关注 Tumeryk，在真诚利落的道谢声中正式结束了本次演讲。 - **离线演示与互动对接**：由于台上分享时间受限，Rohit Valia 留在现场为有意向的嘉宾提供面对面的产品实操演示与咨询。 - **扫码预约机制**：参会者可通过扫描现场幻灯片上的二维码，快速预约 Tumeryk 团队的线上定制演示。 - **致谢与跟进**：Rohit 对全场听众致以谢意，鼓励企业客户进一步了解 Tumeryk 如何帮助其打破“反向信息悖论”并重筑 AI 信任边界。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=nqv-1YorfYE&t=4s) 价值并未沉淀在知识的所有者手中，反而单向流向了模型厂商。 — Rohit Valia > [00:04](https://www.youtube.com/watch?v=nqv-1YorfYE&t=4s) 提示词（Prompts）本质上是企业的核心知识产权，蕴含着最深层的业务逻辑。 — Rohit Valia > [00:04](https://www.youtube.com/watch?v=nqv-1YorfYE&t=4s) 这些在消费级场景中被视作偶发漏洞的事故，在严苛的企业生产环境中均属于零容忍事故。 — Rohit Valia > [03:27](https://www.youtube.com/watch?v=nqv-1YorfYE&t=207s) 任何提示词都不会离开组织内部 — Rohit Valia > [03:27](https://www.youtube.com/watch?v=nqv-1YorfYE&t=207s) AI Trust Score™（AI 信任评分）被定义为针对“虚拟人类”（AI Agent 与大模型）的动态信用评级标准。 — Rohit Valia > [05:36](https://www.youtube.com/watch?v=nqv-1YorfYE&t=336s) 业界厂商声称模型发生“逃逸或失控”是不可原谅的技术失职。 — Rohit Valia > [05:36](https://www.youtube.com/watch?v=nqv-1YorfYE&t=336s) 企业必须部署严格的网关守门人机制；一旦未经受控的模型向外部其他银行或医疗系统发起攻击，后果将是灾难性的。 — Rohit Valia > [07:38](https://www.youtube.com/watch?v=nqv-1YorfYE&t=458s) 员工到底在让 AI 写什么代码 — Rohit Valia > [12:39](https://www.youtube.com/watch?v=nqv-1YorfYE&t=759s) we might not have time for demo — Rohit Valia > [12:39](https://www.youtube.com/watch?v=nqv-1YorfYE&t=759s) just scan this QR code we'll be happy to set up a demo for you — Rohit Valia > [12:39](https://www.youtube.com/watch?v=nqv-1YorfYE&t=759s) I appreciate you all coming — Rohit Valia ## Mentioned Rohit Valia, Satya Nadella, Tumeryk, Ai4, OpenAI, Google, Anthropic, Hugging Face, Cloud Security Alliance (CSA), Gartner, FICO, OWASP, Splunk, Datadog, AI Trust Score™, Kubernetes, MCP (Model Context Protocol), RAG, ChatGPT, Gemini, Grok, Lovable, GDPR, HIPAA, ISO 42001, EU AI Act, 影子 AI (Shadow AI), DLP (数据防泄露) ## Video description Rohit Valia - CEO & Co-founder, Tumeryk Satya Nadella calls it the Reverse Information Paradox: enterprises pay for AI twice, in dollars, and in the proprietary knowledge every prompt and agent workflow reveals to models they don't control. When learning flows one way, value accrues to whoever owns the model, not the knowledge. This live demo shows how runtime governance rebuilds the trust boundary: guardrails intercepting risky prompts and agent actions, risk scoring quantifying exposure, and sovereign AI keeping knowledge in-house. [Source](https://www.youtube.com/watch?v=nqv-1YorfYE)

Source quotes

[0:04] Rohit Valia: 价值并未沉淀在知识的所有者手中，反而单向流向了模型厂商。 [0:04] Rohit Valia: 提示词（Prompts）本质上是企业的核心知识产权，蕴含着最深层的业务逻辑。 [0:04] Rohit Valia: 这些在消费级场景中被视作偶发漏洞的事故，在严苛的企业生产环境中均属于零容忍事故。 [3:27] Rohit Valia: 任何提示词都不会离开组织内部 [3:27] Rohit Valia: AI Trust Score™（AI 信任评分）被定义为针对“虚拟人类”（AI Agent 与大模型）的动态信用评级标准。 [5:36] Rohit Valia: 业界厂商声称模型发生“逃逸或失控”是不可原谅的技术失职。 [5:36] Rohit Valia: 企业必须部署严格的网关守门人机制；一旦未经受控的模型向外部其他银行或医疗系统发起攻击，后果将是灾难性的。 [7:38] Rohit Valia: 员工到底在让 AI 写什么代码 [12:39] Rohit Valia: we might not have time for demo [12:39] Rohit Valia: just scan this QR code we'll be happy to set up a demo for you [12:39] Rohit Valia: I appreciate you all coming

Source quotes

Satya Nadella: enterprises pay for AI twice, in dollars, and in the proprietary knowledge every prompt and agent workflow reveals to models they don't control. Rohit Valia: When learning flows one way, value accrues to whoever owns the model, not the knowledge. [0:04] Rohit Valia: prompts are your intellectual property [2:31] Google Gemini: please die [2:31] Rohit Valia: virtual humans [4:31] escaped [7:36] Rohit Valia: the deepest thoughts of your employees Satya Nadella: Reverse Information Paradox

Original transcript

[0:04] So uh thanks for being here and in fact the  previous talk was very useful to set the stage for [0:10] uh what I'm going to talk about which is the  security uh paradigm for enterprises to be able [0:16] to use the um AI in in a safe and secure manner.  And so before I uh go into how we solve that [0:26] problem uh in the age of AI essentially what we  see is that you are and this was the hugging face [0:34] example that the previous speaker just uh talked  about where by the hugging face team was unable to [0:42] use open AI's model. So essentially the provider  refused to give them access to do the kinds of [0:47] things they wanted to do. And so essentially,  not only are you paying these providers, [0:53] sometimes they can even uh knock off your access  to their models and you're also paying for them [0:59] with the data that you're sending which they  can use for training. So so the whole notion of [1:05] uh being able to use external public models uh  kind of um is not very safe for enterprises to [1:13] use and I'll kind of go into a little bit more  details about that. So when you think about it, [1:19] prompts are your intellectual property and  by using a public service, you're essentially [1:26] uh providing this either to external model  providers. So they are either able to [1:32] uh gain deep insight into your business. Uh they  may train their own models on this data. Uh they [1:39] can use it to enter different markets and they  might be competing against you using your data [1:45] essentially. And so and also it's reverse  engineering some of the work that you're doing. [1:51] Now on the other hand, there are some SAS security  tools that you might be using uh which you use to [1:57] protect and they're very easy to use because they  are just online and you're going to um give you [2:02] they give them your prompts and they'll tell you  if they're safe or not. Well, in that doing so, [2:08] you're again sending your data to those models  externally and they will be using those to further [2:13] train their models which obviously helps everyone.  But in the other hand, your intellectual property, [2:18] you might be at the risk of regulations like  GDPR or you might be uh violating HIPPA in some [2:25] cases. So the approach of essentially giving  your prompts out to third party services is [2:32] uh not very safe for most regulated enterprise and  so there are already risks that we are seeing. So [2:38] there's risks of bias, there's cyber risk, there's  accountability risks in an enterprise and there's [2:45] explanability risks that exist with AI as it  is. And there have been various many real world [2:52] incidents even before the hugging face incident  caught um the attention of the people because [2:57] open AI was the culprit in that case. But there  have been many other instances where lovable was [3:02] weaponized, Grock was used to spread you know  all kinds of disinformation. Uh codebase was [3:09] wiped out. Uh Gemini told a student to please die.  So there there are various incidences which have [3:15] kind of we've kind of accepted as okay they  happened but really in an enterprise setting [3:20] that is not an acceptable answer. So what Tumeric  does is we provide a secure governance of the AI [3:28] uh infrastructure at scale. Essentially providing  you a Kubernetes uh infrastructure inside your [3:36] organization. So no prompts leave and if you use  private models we also provide the hardening and [3:42] the red teaming and then being able to provide  the gateways around the model access MCP and rag [3:50] in a single platform which allows you to also  score every prompt and response in real time. [3:57] So I was at FICO building the platform for fraud  detection and for other use cases around financial [4:05] services and real time analysis is very needed and  so we use similar technology to build the AI trust [4:13] score core which is now the global standard with  the cloud security alliance to be able to measure [4:18] trust in models and agents and if you go to  turmeric.com and I'll have a link later um we are [4:25] also the Gartner cool vendor in this category for  AI cyber security governance. And so essentially [4:31] what we do is think of the AI trust score as the  trust score for your virtual humans just like the [4:38] credit score is a score for various humans that  are applying for credit. And what's common here [4:45] is that it's a non-deterministic problem. So we  are not saying like the score is uh the beall [4:51] and endall of all scores but it is a guidance as  to which is more safe. Like if you had to loan to [4:57] a million users, which credit scores would you  rather loan them to? Right? And so similarly, [5:03] if you have to pick agents and if you have to pick  models, you should uh evaluate obviously different [5:08] properties, but the trust cores provide you a way  for these agentto agent interactions to happen. [5:15] So our platform implements the full agent-to-agent  protocol, passes trust scores between agents, [5:21] authenticates and verifies the identity of the  human that actually initiates the call against [5:28] these agents. So um it's really designed for the  agentic use case and um the trust score uses these [5:36] six different categories and pillars. So I wrote  the paper with the cloud security alliance on how [5:42] to evaluate trust in these uh AI agents and AI  models. We've been doing this for over two years. [5:49] We have scores for tens of these frontier models,  different generations of open AI, different [5:55] generations of Google models etc. which you can go  and check out for free on our website [snorts] uh [6:01] for which we'll have a link later on. And what  we've done is we take these six different pillars [6:07] and then allow you to evaluate trust in these uh  different categories. So you may say I I don't [6:14] care about this particular uh pillar and so you  can evaluate trust in each of these individual [6:21] categories as well. And then the system also  allows you to run like OASP and other security [6:27] uh related um uh probes that can help you deeply  inspect these models for what kinds of falsehoods [6:35] or um uh hallucinations or even prompt injections  that they might fall for as well as jailbreak [6:42] attacks and those are very very easy. So without  so when somebody says like open somebody like open [6:48] AAI says that hey um the model escaped that is  unforgivable because they should have a check they [6:54] should have a gateway which should be looking at  what prompts is their uh system you know sending [7:00] out. So it should not be uh that hard and and for  enterprises that can be catastrophic if your model [7:07] starts to go and attack some other bank's model  or some other healthcare institution. So, so the [7:13] the system needs to be properly regulated end to  end all the way from what access these models are [7:22] going to have as well as what are the different  ways in which you can um enable the users to get [7:30] access to different capabilities and that way you  can trace throughout the life cycle. Um so the way [7:38] we run our system is we have a uh set of uh models  that we deploy inside your organization that look [7:47] at the prompts and responses for everything  that traverses in real time. So it's not like [7:54] we are doing sampling. It's like every prompt and  response is being evaluated for trust along the [8:00] guide uh the guardrails that you configured. And  the idea is that this is your infrastructure. No [8:08] prompt should leave your organization. Prompts  are your uh intellectual property. They are [8:14] the deepest thoughts of your employees. And any  service where you're sending your prompts outside, [8:21] you are losing that intellectual property. And  um the three offerings that we uh uh offer our [8:31] clients is essentially what we call AI posture  intelligence. So with AI posture intelligence, [8:37] what we do is we come and evaluate your  infrastructure using our tools. uh we [8:42] gave you an assessment of all the shadow AI that  might be happening in your organization as well as [8:49] uh establish a catalog of assets that you have and  discover all your assets and then you can assign [8:56] these assets to different teams and once you start  operationalizing whether it's your workforce that [9:03] is accessing models externally like Gemini or chat  GPD um or it's developers so even developers using [9:11] coding agents, you need to monitor their usage.  What are they asking it to code? How are your [9:16] tokens being spent? Every prompt and response  is being monitored in this system. And on the [9:24] agentic governance side, we have a security  gateway for real-time threat mitigation, [9:29] which means we are not just observing, we can  mitigate, we can block a prompt, we can blur an [9:35] image face if it's PII control, like in a HIPPA  setting. we can mask the data in real time as the [9:42] prompt is exiting an environment. So, so those are  all capabilities in the platform and this helps [9:49] you comply with all the regulations. So, at the  end of it, once you put the system operational, [9:55] we'll give a PDF report on a per uh department  on a per engineering team level and at the exact [10:03] management level with whichever framework  you choose. So we will be able to develop [10:08] that based on the data we've collected  as to what are the gaps in your ISO 4201 [10:14] what are the gaps in your EU AI act um you know  capabilities so with that what you want is that [10:22] uh to summarize we start with an assessment a  risk assessment of your environment so you would [10:30] uh inspect it you would run these probes that  can go and find all the skills all the agents [10:36] all the models that you have in your environment.  So you want to first do that piece as as you're [10:42] going through this uh process and then you move on  to the um you know making sure your uh employees [10:50] and workforce is also doing uh is monitored and  your agents are also monitored. So the the trust [11:00] platform is deployed within your organization  and it takes care of your AI risk assessment [11:07] as I mentioned the shadow AI detection agent  governance the data loss prevention uh all the [11:15] compliance reporting that you would need for  your AI agents that are essentially governed. [11:21] So you have an asset catalog and for every asset  we'll be able to show you a trust score for that [11:28] asset that AI asset whether it's a model it's an  agent and in future we adding MCP server so this [11:35] is the work that we did with the cloud security  alliance for which I wrote the paper and then you [11:40] get incident response so you can integrate  this with your enterprise tools you don't [11:44] have to use our dashboards if you use splunk for  logging then you can use that if you're using you [11:51] data dog for monitoring. You can send those  incidents or uh alerts over there and then [11:57] you're getting your workforce security as well  because um your workforce may be uploading very [12:02] sensitive data into external services without you  knowing. So we are working with a medical health [12:08] science center where they wouldn't want their  employees to by mistake or even intentionally [12:14] upload patient information. That's a HIPPA  violation. So they have actually locked down [12:19] the use of private accounts by uh requesting  open AI and anthropic in that they should not [12:26] allow private accounts from their organization  to be created and everything that is being used [12:32] is governed and has to go through the tumeric  gateway as a way to evaluate trust in these [12:39] uh systems. And uh we might not have time for  demo but I'm available. So if anybody wants to [12:46] uh you know get a demo I'll be around and we'll  do a demo for you or certainly you know if you [12:53] um just scan this QR code we'll be happy to  set up a demo for you. So um I think we're uh [13:01] done with my time. So with that I appreciate you  all coming and certainly look us up. Thank you.