AgentSnare traps AI pentest agents in decoy systems

AgentSnare fooled AI pentest agents into filing false victory reports 90% of the time, per the arXiv paper. That matters for anyone building autonomous red-team tools: a confident, evidence-backed report of a fake win poisons an operation's results before it ends.

Older honeypots plant static honeytokens and wait. AgentSnare instead watches an agent's interaction history and builds new decoy artifacts to match whatever it just tried. Each validated fake gets folded into a consistent environment as the session continues.

Across 45 attacker-CVE pairs on 15 CVE-Bench apps, three attacker models chased decoy artifacts instead of the real target, and none landed an exploit. The paper reports it absorbed 46.8% of their tool calls and kept 55.9% of their post-entry actions inside the fake system.

The paper doesn't say whether AgentSnare holds up against an attacker model trained to notice it's being fed decoys. That's the test that would settle it.

If decoy systems like this hold up against models built to detect them, report-poisoning rate becomes the metric to watch. Not block rate. The arXiv paper is new as of July 30, 2026.