A trace study of AI agents repairing software vulnerabilities found that most of 95 confirmed “silent failures” were linked to a missed security requirement or an inadequate defense. The patches passed the available functional checks but still contained a relevant weakness, the researchers report.
The team developed Security Awareness Gap Evaluation (SAGE), which scores each logged reasoning turn for security awareness and lines it up with saved versions of the code. It evaluated 95 manually confirmed failures from 3,684 valid traces generated by six open-source agent frameworks and six base models on tasks drawn from SecurityEval and CVEfixes. The authors’ June study categorized silent failures; this paper focuses on where a failure first became visible in an agent’s trace.
SAGE assigned an origin in 93 cases: 51 were classed as an unaddressed requirement, 37 as an inadequate defense and five as a defect in the code change; two cases were undetermined. When an agent introduced vulnerable code, the assigned origin came before the edit in 14 of 19 cases; the other five were single-message runs with no earlier recorded turn.
SAGE’s rule identifies the earliest trace turn that appears to diverge from the task’s security intent; it does not establish that the turn caused the vulnerability. Repeated scoring and a second judge agreed on failure category more often than the exact turn. Human checks assessed code history and edit location, not whether SAGE’s reasoning-origin label was correct. The researchers did not test whether adding a security check at the identified point would prevent such failures.
Photo: Markus Spiske/Pexels, used under the Pexels License. Illustrative stock image, not a study screenshot.