AI Trends

OpenAI Opens Its Internal AI Systems to On-Site Audits by Four Independent Organizations for the First Time

The DecoderUpdated Nov 19 02:19

AI summary

For the first time, OpenAI is allowing Apollo Research, METR, Transluce, and Goodfire to directly audit its internal AI systems at its offices, investigating safety risks and model behavior. Depending on their assignments, the organizations may access unreleased models, source code, training data, and monitoring systems, and hold confidential discussions with developers. An independent Safety and Security Committee will oversee the process and receive unredacted reports. The auditing organizations may publish their findings independently, but OpenAI may redact confidential material or delay reports so it can address risks first.

Why read this

AI score 73

Four external organizations will gain access to unreleased models, source code, and training data, with the on-site audit framework outlining the specific boundaries and governance processes for safety reviews of frontier models.

Topics

Tags

Body

OpenAI is giving four outside audit organizations direct access to its internal AI systems to investigate safety risks and model behavior on-site. The company says it's working with Apollo Research, METR, Transluce, and Goodfire, as OpenAI announced. The program builds on an earlier on-site review of its monitoring systems and fulfills commitments from the "White House Accord on Super Intelligence."

Depending on their assignment, the organizations get access at OpenAI's offices to unreleased models, source code, training data, and monitoring systems. They can also speak confidentially with developers. Apollo Research is examining whether models secretly work against their developers' intentions. METR is reviewing AI misbehavior in the largest reinforcement learning training run, a method where models learn through rewards. Goodfire is analyzing internal model processes, and Transluce is evaluating post-training safety checks. OpenAI and Transluce are also planning an eight-week follow-up review for early 2027.

An independent safety board oversees the entire process, receives unredacted reports, and mediates disputes. The auditors publish their findings independently, but OpenAI can redact confidential material and delay reports to fix risks first. The company expects some friction early on but is already considering more reviews, including ones focused on company culture and internal processes.

Source

The Decoder