safety security
The Agent's Failure Usually Starts Near the Beginning
A trajectory audit finds contradictions were already present before failure in most reviewed runs, but the systems rarely acted on them.
Summary
A trajectory audit finds contradictions were already present before failure in most reviewed runs, but the systems rarely acted on them.
Researchers read 95 failed multi-agent trajectories in full and annotated where each fault originated, how it propagated, what evidence contradicted it and which control could have reached it. The median fault entered at 18% of the run, usually as a wrong task-state decision or fabrication. In 72% of failures the trajectory already contained evidence against the error; in 88% of those cases no agent acted on it. The audit also found that 45% of one corpus's hand-crafted failures ended in a provider content-filter crash followed by a harness-generated answer, exposing a failure class that responsible-agent labels miss.
Why it matters
A trajectory audit finds contradictions were already present before failure in most reviewed runs, but the systems rarely acted on them.
Limits and context
No additional limitation was separately recorded.
Key claims
A trajectory audit finds contradictions were already present before failure in most reviewed runs, but the systems rarely acted on them.
Evidence: source-2026-11-12-005
Sources
- arXiv:2611.13254v1arXiv / Cornell University · primary research
Corrections
No corrections have been recorded for this story.