Enterprise agent security whitepaper maps 8 vulnerability classes, claims 8.3% injection rate

Kyrannio · x · 2026-09-25

Pokee AI, with Bo Li's teams at UChicago/UIUC, released an "Enterprise Agent and Model Security" whitepaper arguing you must secure the agent trajectory, not just the model.
Three parts:
Maps the enterprise agent attack surface: user objective → untrusted content → planner decision → tool composition → persistent action; risk only visible in full trajectories;
Identifies 8 vulnerability classes that standard application-security tooling misses (models/planners, tools & skills, MCP metadata, etc.);
Proposes a hardened enterprise agent architecture.
Introduces DecodingTrust-Agent, a benchmark (12 domains, shared judge, 6,200 tasks per model) testing Pokee-Isaac 28B v0.1 against five mainstream models:
Indirect prompt-injection success: Pokee-Isaac 8.3% vs 36.7% for runner-up Claude Haiku 4.5; GPT-5.6 Luna 46.0%, Qwen3.5-122B 46.1%, Gemini 3.5 Flash-Lite 49.0%;
Direct attack success 30.5% (lowest), benign task success 85.6% (highest).
Caveat: vendor-built model on the vendor's own benchmark; independent replication needed.

More from coding & agent