AI systems could cover up misbehavior

Research question

Can the evidence viewer itself undermine oversight?

Design & assumptions

An assisted proof of concept tested integrity of the evaluation transcript viewer.

Findings

Displayed evidence could be altered while database records remained intact.

The maintainers patched the issue; no exploitation in actual evaluations was observed.

Limits & counterevidence

A demonstration of a viewer flaw is not evidence of autonomous exploitation frequency. Operational exploit details are excluded.

Source evidence

No structured numerical extraction is available in this reading package.