Non-profit evaluation organizations METR and Redwood Research stated on Thursday that they conducted a pilot review of OpenAI's internal monitoring system. This is the latest attempt by this artificial intelligence company to assess how it can deepen its cooperation with third-party security organizations.
METR and Redwood indicate that this experiment is intended to let OpenAI test what it would be like to provide 'deep access' to its internal infrastructure to external researchers.
According to a statement, between July 21 and August 11, one employee from each of these two organizations was granted access to the OpenAI internal code repository, documentation, and some technical staff resources. The statement mentioned that these employees were always accompanied while on-site and used devices provided by OpenAI, which "never left the premises."
Evaluators have discovered several "new vulnerabilities," including a method that could allow an entity to prevent "hundreds of" recent operations from being monitored. According to the statement, OpenAI informed METR and Redwood that the identified vulnerabilities affecting their production protection measures have been patched.
METR and Redwood indicated that they agreed to publish only one high-level public post regarding the findings, as this pilot was not intended to 'provide strong public transparency for OpenAI's monitoring system.' They submitted a 61-page internal report to OpenAI and stated that they expect to share more details about subsequent trials in the future.
OpenAI said to CNBC: "This collaboration is built on our long-term commitment to pioneering new methods of third-party testing. We are utilizing what we have learned to strengthen our protective measures and to shape the way we will cooperate with independent assessment institutions in the future."
OpenAI CEO Sam Altman promised in September to provide "employee-like access rights" to independent assessment agencies. Previously, Anthropic CEO Dario Amodei proposed this idea as part of his plan to slow down the development pace of AI. Subsequently, Anthropic announced agreements with institutions such as Accenture, METR, and SecureBio. OpenAI stated earlier this month that they are "actively finalizing contracts with third-party security assessment agencies."
U.S. President Donald Trump, who adopts a relatively lenient AI regulatory stance, also expressed support for third-party evaluations. In a voluntary agreement proposed at the White House, he encouraged AI companies to "cooperate with independent external auditors or evaluators." The Amodei of Anthropic signed this agreement, as did the executives of Google, Meta, OpenAI, SpaceX, and Nvidia.
Independent assessment agencies have always been in a relatively quiet corner of the trillion-dollar AI industry. Therefore, experts say that the outside world is still concerned about how these agencies will obtain funding and what kind of reporting frameworks they will ultimately adopt.
METR and Redwood stated on Thursday that during their three-week pilot, they conducted stress tests on the mismatch monitoring system within OpenAI. In the field of AI, alignment refers to the concept used by researchers to describe whether a model aligns with human interests and values.
The evaluators assessed the monitoring system as of July 21st and prior to that date, and noted that OpenAI was in the process of expanding the system during the trial period.
View:Former OpenAI executive raises funds for a defense startup valued at $1.5 billion
