Did the attack work? Auditing AI-agent security benchmark verdicts
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
This Zenodo preprint (v1.2, by TRUSCOR Private Limited's author, Nov 17 2026) audits the automated verdicts of AgentDojo, a benchmark for prompt-injection attacks on AI agents.