External Evaluators Say They Found OpenAI Monitoring Flaws
By @sharedot · · 6 pages
- AI
- AI Safety
- Openai
- Third Party Evaluation
Evaluators METR and Redwood Research say a three-week review found novel vulnerabilities in OpenAI's monitoring system, since patched.
What happened: the first deep-access pilot
Nonprofit evaluators METR and Redwood Research announced Thursday that they conducted a pilot review of an internal monitoring system at OpenAI, designed to test what it would look like to give external researchers "a deep level of access" to the company's internal infrastructure.
The exercise ran from July 21 to August 11, with one employee from each organization granted access to OpenAI's internal codebase, documentation and some technical staffers.
The researchers were chaperoned on site and used OpenAI-issued devices that never left the premises, according to the evaluators' release.
Why it matters: flaws in the safety net
The surprise is that the very system meant to catch misaligned AI behavior had holes an outside team could find.
The evaluators say they discovered several "novel vulnerabilities," including one exploit that would allow agents to prevent "hundreds" of recent actions from being monitored — effectively letting a model operate while evading oversight.
OpenAI told METR and Redwood that it has patched the identified vulnerabilities affecting its production safeguards.
According to CNBC, an OpenAI spokesperson framed the collaboration as part of the company's commitment to "pioneering new approaches to third-party testing" and strengthening future engagements with independent evaluators.
The evidence, and what stays private
METR and Redwood stress-tested the misalignment monitoring system as it existed on and before July 21, noting that OpenAI was expanding the system during the exercise itself.
They submitted a 61-page internal report to OpenAI but agreed to publish only a high-level public post, explaining that the pilot was not designed to "provide robust public transparency about OpenAI's monitoring system." They said they expect to share more details about future exercises.
Stakes: third-party auditing goes mainstream
The pilot lands as independent evaluation shifts from a quiet corner of the AI industry toward formal infrastructure.
Anthropic has since announced agreements with organizations including Accenture, METR and SecureBio, and OpenAI said earlier this month that it is "actively finalizing contracts with third-party safety assessors."
President Trump has also endorsed the approach, encouraging AI companies to "partner with an independent external auditor or evaluator" under a voluntary accord signed by executives at Anthropic, Google, Meta, OpenAI, SpaceX and Nvidia.