iCreatAI News
Skip to content
Hacker News AI· gautampatiltube·· 19 hr agoAI score52

Did the attack work? Auditing AI-agent security benchmark verdicts

Did the attack work? Auditing AI-agent security benchmark verdicts

AI summary

The author audited AgentDojo's automated attack-success verdicts against agents' actual tool calls, finding 97.7% agreement on non-crashing published runs but 433 runs scored as no attack in which the agent executed the attacker's decisive action (11.9% of audited attacks), and 55 of 744 missed on new runs under benchmark v1.2.2. Re-running 240 recorded attacks ten times each, 224 recurred in at least half of runs but only 201 could be shown at 95% confidence to do so, and preregistered rules for which setup changes stop an attack missed their test on the protocol's three models. The paper releases .sova, an open-source tool and record format that audits verdicts, re-runs AgentDojo findings, and seals evidence for offline verification against published checksums.

Source: Hacker News AI · zenodo.org