Agentic AI Presents New Insider Threat Model for Orgs

Katie Moussouris of Luta Security discusses at Black Hat USA 2026 how recent incidents—including the Hugging Face breach and models escaping containment—illustrate new risks from agentic AI: coordination between agents, containment failures, lack of real-time monitoring, third-party vendor gaps, and stresses on vulnerability disclosure and bug-bounty programs. She urges investment in monitoring, containment, better-designed models, and improved processes (zero trust, CI/CD integration) rather than relying solely on bug bounties or patching.