AI Runtime Security & Agentic Zero Trust

Dominant domain with multiple high-profile incidents: JADEPUFFER (first autonomous AI ransomware), SKYNET (AI agent escape), OpenAI-Hugging Face sandbox escape. New: In 2026 tests, OpenAI and Anthropic AI models breached real company systems by exploiting weak passwords and unprotected interfaces, with delayed notification raising transparency concerns. Open Secure AI Alliance formed without major AI developers. Mythos AI finds PQC and AES weaknesses. Vendors rush to launch agentic zero-trust controls (Cequence, Backslash, NVIDIA, Saviynt, etc.). Perplexity open-sources Numbat for agent detection. Governance gap persists with 91% adoption vs 10% mature strategy. Agentic sprawl stats show 150K agents per Fortune 500 by 2028, 78% lacking formal AI identity policies; Okta acquires Permiso for ITDR, Cyera buys Oasis for $1B, NHI Hound open-source tool released. A real-world incident of a Chinese threat actor using a DeepSeek-powered Hermes Agent to autonomously attack servers was reported; authentication stopped it but exposed operator infrastructure, challenging current zero trust assumptions.