Pokee and UChicago/UIUC researchers map 8 enterprise AI agent vulnerability classes in new whitepaper
Kyrannio · x · 2026-09-25
Pokee AI released a whitepaper on Enterprise Agent and Model Security, co-authored with Bo Li's team at UIUC and the University of Chicago. It maps the enterprise AI attack surface, identifies eight vulnerability classes that standard application-security tooling isn't built to detect, and proposes a hardened enterprise architecture. The team also benchmarked Pokee-Isaac 28B v0.1 (Preview) on DecodingTrust-Agent against GPT-5.6 Luna, Claude Haiku 4.5, Gemini 3.5 Flash-Lite, Qwen3.5-122B and Nemotron 3 Super 120B based on those vulnerabilities.
Related event: Pokee, UChicago and UIUC Release Enterprise Agent Security Whitepaper(2 posts)→
More from Safety
- GitHub Security Lab open-sources an AI fuzzing pipeline that writes and tunes its own harnesses — GitHub Blog AI/ML · 2026-09-25
- Apple launches Reference Image to prove iPhone 18 Pro photos aren't AI slop — ChuckDBrooks · 2026-09-25
- Scoop: NSA is spending billions this year testing AI models, per classified estimate — peterwildeford · 2026-09-25
- As bots hammer websites, IP-level restrictions are closing the open internet — mobileraj · 2026-09-25
- White House asks OpenAI and Anthropic to withhold new models from UK testers — Polymarket · 2026-09-25
- Shanghai AI Lab and universities unveil SHE and SafeEvolve to secure agents via full execution trajectories — jiqizhixin · 2026-09-25