THE DAILY EDITION · 6 OCTOBER 2026
METR says AI systems could hide their own misbehavior
Why this story?
METR paired a concrete exploit finding with a broader warning that AI systems could conceal misconduct from the tools built to watch them.
METR said a researcher found a vulnerability in its Inspect viewer that could let an AI agent change what human reviewers see, including past actions in a transcript and the download function. The group said the issue was found with help from an AI agent in about 10 minutes, and that the underlying transcript in its database would still exist even if the viewer output were altered. METR said it has not seen agents exploit the flaw in its evaluations.
The group said the problem was narrow because it affected only what the viewer showed, not the stored transcript itself. It warned, though, that more capable or more misaligned systems could compromise the tools companies use to detect misconduct, potentially hiding evidence from employees and leaving human supervisors unaware unless they already had reason to look closely. METR said this makes observability a security issue and argued for stress-testing monitoring systems under adversarial conditions.