Skip to content
Brief IA·Indexed by this site · Originally published· Governance →· Its place in the scan record →SelectedAttention59

Four organizations enter OpenAI for the first time to audit its internal AI systems

Quatre organismes auditent les systèmes internes d'OpenAI | Brief IA

AI summary

For the first time, OpenAI hosted four independent organizations—Goodfire, Transluce, METR and Apollo Research—at its own offices to directly audit the safety and model behavior of its internal AI systems. The scope of the audit depends on the task and includes unreleased models, source code, training data and monitoring facilities, and confidential interviews with developers are also possible. An independent safety committee is responsible for overall oversight, reading the full reports and mediating in case of disagreement; the auditing organizations publish their results autonomously, but OpenAI may remove confidential content or delay publication so that identified risks can be addressed first. In terms of division of labor, Goodfire studies the models' internal mechanisms, Transluce reviews post-training safety controls and plans a new eight-week review with OpenAI in early 2027, METR looks for misconduct during the largest-scale reinforcement learning training, and Apollo Research examines whether models will act covertly against developers' intentions. The company expects friction during the startup phase and is considering a separate review of corporate culture and internal processes.

Why it was selected

Four external institutions were permitted to enter OpenAI to view unreleased models and training data, but the company retains the power to delay release; this division of authority is key to understanding the independence of third-party audits.

Read originalbriefia.fr