METR, a nonprofit rating agency, and Redwood Research announced Thursday that they conducted a pilot review of the internal monitoring system. OpenAIthe artificial intelligence company’s latest effort to determine how to deepen its collaboration with third-party safety organizations.
METR and Redwood said the exercise was designed to test what it would be like for OpenAI to give outside researchers “deep-level access” to its internal infrastructure.
From July 21st to August 11th, employees at each organization had access to OpenAI’s internal codebase, documentation, and some technical staff. The employee was escorted while on-site and used an OpenAI-issued device that “never left the premises,” according to the release.
Evaluators discovered several “new vulnerabilities,” including an exploit that allows the agent to prevent monitoring of “hundreds” of recent actions. According to the release, OpenAI told METR and Redwood that it has patched identified vulnerabilities that impact security measures in production environments.
METR and Redwood said they agreed to publish only high-level public posts about the findings because the pilot was not designed to “provide robust public transparency regarding OpenAI’s surveillance system.” They shared a 61-page internal report with OpenAI and said they plan to release more details about future exercises.
“This partnership builds on our long-standing commitment to pioneering new approaches to third-party testing, and we are leveraging what we have learned to strengthen our safeguards and shape future collaborations with independent evaluators,” an OpenAI spokesperson told CNBC in a statement.
OpenAI CEO Sam Altman pledged in September to give independent evaluators “employee-like access” to the company, after Anthropic CEO Dario Amodei pitched the idea as part of a plan to slow the pace of AI development. Since then, Anthropic has announced deals with organizations including: AccentureMETR, SecureBio. OpenAI said earlier this month that it is “actively entering into agreements with third-party safety reviewers.”
President Donald Trump, who has taken a moderate approach to AI regulation, has also expressed support for third-party evaluations. As part of a voluntary agreement he presented at the White House, he encouraged AI companies to “partner with independent external auditors or evaluators.” Anthropic’s Amodei signed the agreement, along with executives from Google, Meta, OpenAI, SpaceX, and Nvidia.
Because independent evaluators have historically been a relatively quiet corner of the multitrillion-dollar AI industry, experts told CNBC that questions remain about how these entities will be funded and what their reporting structure will ultimately look like.
METR and Redwood announced Thursday that they stress tested OpenAI’s internal misalignment monitoring system during a three-week pilot. In AI, alignment is how researchers describe whether a model behaves in line with human interests and values.
The evaluators evaluated the surveillance system that existed before July 21 and noted that OpenAI had expanded the system during the exercise.
WATCH: Former OpenAI top employee raises $1.5 billion worth of funding for defense startup
