Evaluators say 'vulnerabilities' found with OpenAI monitoring system - UPI.com

Oct. 22 (UPI) -- Two nonprofit organizations that evaluate AI safety said Thursday that they found weaknesses in an OpenAI monitoring system. The findings came from a trial project testing how the company will work with outside safety organizations in the future.

According to METR and Redwood Research, their trial review of OpenAI's system tested what it would be like for the company to give outside evaluators deep access to its internal systems.

The organizations said one employee from METR and one from Redwood had access to OpenAI's code, documents and some technical staff from July 21 to Aug. 11. OpenAI employees supervised the two researchers while they used company devices, which never left the company's buildings.

The evaluators said they found several new weaknesses. One of them could allow AI agents to keep hundreds of recent actions from being tracked by the monitoring system.

The researchers said OpenAI told them it had fixed the weaknesses that affected its safeguards in its live systems.

METR and Redwood said they gave OpenAI a 61-page internal report. They added that they expect to share more details about their testing activities in the future.

"This collaboration builds on our longstanding commitment to pioneering new approaches to third-party testing, and we're using what we learned to strengthen our safeguards and shape future collaborations with independent evaluators," an OpenAI spokesperson said in a statement to CNBC.

Last month, OpenAI CEO Sam Altman promised to give independent evaluators access to the company similar to that of its employees. He made the pledge after Anthropic CEO Dario Amodei proposed a similar approach as part of his plan to slow AI development.

Anthropic has since announced agreements with Accenture, METR and SecureBio.

President Donald Trump has supported outside reviews of AI safety. He has encouraged AI companies to "partner with an independent external auditor or evaluator." OpenAI, Anthropic, Google, SpaceX and Nvidia signed a voluntary agreement that Trump presented at the White House last month.