Go to PegaPoll
PEGAPOLL · NEWS

OpenAI grants four independent auditors access to internal AI systems

The Decoder (AI) · 2026-11-18
🤖 AI-generated content — The title and summary were produced automatically by artificial intelligence, without human editorial review.

OpenAI has opened its internal AI systems to four independent audit organizations—Apollo Research, METR, Transluce and Goodfire—in an effort to examine model behavior and identify safety risks before they become harder to address. The initiative is the company’s first broad on-site review of this kind and follows an earlier assessment of its monitoring systems. OpenAI says it also fulfills commitments made under the “White House Accord on Super Intelligence.”

Depending on their assignments, the auditors can examine unreleased models, source code, training data and monitoring systems at OpenAI offices, while speaking confidentially with developers. Apollo Research is investigating whether models might secretly pursue goals that conflict with their developers’ intentions. METR is reviewing possible misbehavior during the largest reinforcement-learning training run, in which models learn through reward signals. Goodfire is studying models’ internal processes, while Transluce is evaluating post-training safety checks. OpenAI and Transluce plan an additional eight-week review in early 2027.

An independent safety board oversees the program, receives unredacted reports and mediates disputes. The auditors are expected to publish their findings independently, but OpenAI may redact confidential information or postpone publication while it works to mitigate identified risks. That arrangement gives the reviews access to unusually sensitive material, while leaving open questions about how freely critical findings can be released.

Continue in the app — vote & join in ➔
Source: The Decoder (AI) · via ahirlevel.hu

Related news