Agent Governance & Identity Security
Key Questions
What security incidents are occurring with AI agents?
Half of enterprises have experienced AI agent security incidents. The first documented agentic ransomware attack was reported by Sysdig involving Jadepuffer.
What new governance frameworks and tools have launched?
Microsoft released Zero Trust guidance for AI agents using Entra, Sentinel, and Purview. Keeper Security introduced endpoint-level agentic AI governance with OS-level observation.
How are organizations addressing AI agent identities?
78% of organizations lack formal policies for creating or removing AI agent identities. This creates major compliance gaps, especially in regulated industries.
What benchmarks reveal gaps in AI safety?
A child-safety benchmark showed frontier models failing up to 34% on non-explicit risks such as grooming and impersonation. This highlights critical gaps in current safety evaluations.
Which standards bodies are updating agent security guidance?
OWASP opened its Agentic Skills Top 10 for public review. CSA released an addendum on securing agentic AI and NIST published its Cyber AI Profile.
What attack surfaces are associated with MCP?
MCP Security maps vulnerabilities across Host, Client, and Server components. Microsoft has flagged tool description poisoning as a key risk in MCP implementations.
How are vendors enhancing runtime agent security?
Codenotary launched AgentMon 3, which processes 5M daily interactions and reduces policies by 80%. Reco and Exabeam expanded agent detection and behavioral monitoring capabilities.
What practical resources support safe AI agent adoption?
An AI Agent Incident Response playbook outlines a seven-step containment process. Playwright provides enterprise containment strategies for browser-based agents.
Fable 5 restoration confirmed with rapid community adoption. NSA MCP security guidance. New product launches: WitnessAI Agentic Control, Tigera Lynx, Legit Security AppSec agents, Pramaana Labs formal verification, Commvault ResOps, F5 AI Security Platform, Okta XAA ecosystem, Reco Agent Security, Governor Agents framework, Zero Trust for AI agents guidance. Microsoft flags MCP tool description poisoning. Exabeam expands agent detection. Agentic Commerce introduces KYA concept. Practical checklists for safe AI adoption. SR 26-2 excludes gen AI from model risk definition. First documented agentic ransomware attack (Sysdig Jadepuffer). NIST Cyber AI Profile. Online safety monitoring for LLMs. Industry shift to delegation design. New: Omnigent, AI-Infra-Guard, Agent Gateways, practical framework, third-party model risk guide, BlueVoyant, Fable 5 restoration timeline. Zscaler IPI traps confirm architectural vulnerability. Vera framework achieves 93.9% attack success rate. Phantom squatting: 250k domains across 913 brands. Execution layer security critical. Gurucul behavioral AI, Radware compliance reporting. New: CSA addendum on securing agentic AI. New: OWASP Agentic Skills Top 10 open for public review. New: Claude Cowork goes mobile. New: Hybrid model architecture confirmed as mainstream. New: Keeper Security introduces endpoint-level agentic AI governance with OS-level process observation. New: Codenotary launches AgentMon 3 adaptive runtime security (5M daily interactions, 80% policy reduction). New: Microsoft publishes Zero Trust for AI agents guide with Entra, Sentinel, Purview. New: Child-safety benchmark reveals frontier models fail up to 34% on non-explicit risks (grooming, impersonation) – critical gap in AI safety evals. New: Abrigo launches agentic AI platform for lending with 40% labor reduction, human-in-the-loop governance. New: Secret Network cites AI exploit risks in proposed Arbitrum move – ties agent security to blockchain migration. New: AI Agent Incident Response playbook – seven-step process for containment and investigation. New: AI Security Executive Order 14409 – light-touch on model licensing, aggressive on federal AI security posture, explicit mention of AI agents in computer-crime enforcement. New: 78% of orgs have no formal policies for AI agent identities – massive compliance gap for regulated industries. New: Anthropic collaborates on GRAM – modularizing dual-use capabilities like virology, practical approach to capability control. New: MCP Security deep-dive – maps attack surface across Host, Client, Server components. New: Playwright MCP Security guide – enterprise containment strategies for browser-based agents. New: LLM language-dependent behavior – models produce inconsistent outputs based on language, critical for global deployments.