Even the tools used to review agents need security testing
METR says it has not observed agents exploiting this vulnerability in its evaluations. This researcher-directed test is not a newly observed autonomous escape or production incident, and it does not change Agent watch classifications.
What was claimed?
METR described a controlled proof of concept that could alter what an Inspect transcript viewer displayed while leaving stored transcripts intact. The maintainer recorded a fix on June 23; METR published its analysis on October 6.
What does the evidence support?
Independent evidence, scope limited
Logs and the interface displaying them are different layers of evidence. Monitoring tools need protection from untrusted agent output.
What are the limits?
METR says it has not observed agents exploiting this vulnerability in its evaluations. This researcher-directed test is not a newly observed autonomous escape or production incident, and it does not change Agent watch classifications.
Claim or evidence date: 2026-10-06. Last verified: .
Colors classify evidence; they do not rank truthfulness. Explore hype versus reality →