What Standard Evaluations Hide About Detectors of Hijacked LLM Agents: Rare Attacks, Unseen Tools and Benchmark Transfer

Summary

This research audits eleven evaluations of detectors that flag hijacked LLM agents from their tool-call traces and finds attack rates of 16% to 86% with no variation tested. On AgentDrift and AgentDojo, a detector with AUROC 0.879 reaches precision 0.076 at a 2% attack rate, and detector rankings reverse as tool coverage falls below about 50%. Detectors trained on AgentDrift rank AgentDojo traces at chance, while the AlignmentCheck monitor reaches AUROC 0.930 but would still mostly raise false alerts at that attack rate.

Related items

• MediumGHSA-j656-3hf2-fvjc: DBHub: MySQL and MariaDB comment interpretation bypasses read-only SQL validationSimilar attack · GitHub Advisory Database
• MediumCritical Langflow Flaws Could Let Unauthenticated Attackers Run Code on ServersSimilar attack · The Hacker News
• LowChainalysis Discloses Data Breach Following AI-Driven CyberattackSimilar attack · SecurityWeek
• CriticalBeta9 is vulnerable to Host Filesystem Exposure via Volume Path TraversalSimilar attack · JFrog Security Research (Vulnerabilities)
• CriticalBeta9 is vulnerable to a Build Worker Sandbox Escape via Chroot-only Buildah IsolationSimilar attack · JFrog Security Research (Vulnerabilities)