Evaluators find vulnerabilities in OpenAI monitoring system — CNBC

Nonprofit organizations METR and Redwood Research identified several new vulnerabilities during a pilot review of OpenAI’s internal monitoring system. One of them could have allowed agents to make monitoring of the last “hundreds” of actions impossible, CNBC reports.

The review lasted three weeks

The pilot ran from July 21 to August 11. One employee each from METR and Redwood Research received access to OpenAI’s internal source code, documentation, and some of the company’s technical specialists. The researchers worked on devices provided by OpenAI that did not leave the company’s premises, while OpenAI employees accompanied their on-site work.

The review was intended to show what providing outside researchers with deep access to the company’s internal infrastructure could look like. METR and Redwood Research tested the internal system for monitoring model misalignment in the version that existed as of July 21 or earlier. They also noted that OpenAI was expanding this system during the testing.

More current news is available on the UA.News Telegram channel Telegram.

OpenAI said it fixed the issues

According to the evaluators, OpenAI said it had addressed the identified vulnerabilities that affected the company’s safeguards in the production environment. METR and Redwood Research submitted a 61-page internal report to the company but agreed to publish only a general post. They explained that the pilot was not intended to provide full public transparency about OpenAI’s monitoring system.

An OpenAI spokesperson told CNBC that the collaboration is part of the company’s approach to independent testing and is intended to help strengthen safeguards and shape future engagement with external evaluators. In September, OpenAI CEO Sam Altman promised to give independent evaluators access similar to that of company employees amid growing concerns about artificial intelligence safety.