Tech Meridian ← LIVE FEED
PROMY MERIDIAN RU

NEWS · COMPANIES · #3834

OpenAI gives four independent auditors on-site access to internal AI systems

OpenAI is allowing four outside audit organizations—Apollo Research, METR, Transluce, and Goodfire—on-site access to its offices and internal AI systems, including unreleased models, source code, training data, and monitoring tools to investigate safety risks and model behavior. Assignments include examining model misalignment, a large reinforcement-learning training run, internal model processes, and post-training safety checks; an independent safety board will oversee the process, auditors can publish findings independently, and OpenAI may redact or delay reports to mitigate risks; OpenAI and Transluce also plan an eight-week follow-up review in early 2027.

KEY POINTS

  1. OpenAI is allowing four outside audit organizations—Apollo Research, METR, Transluce, and Goodfire—on-site access to its offices and internal AI systems, including unreleased models, source code, training data, and monitoring tools to investigate safety risks and model behavior.
  2. Assignments include examining model misalignment, a large reinforcement-learning training run, internal model processes, and post-training safety checks; an independent safety board will oversee the process, auditors can publish findings independently, and OpenAI may redact or delay reports to mitigate risks; OpenAI and Transluce also plan an eight-week follow-up review in early 2027.
  3. This is a notable expansion of external oversight into OpenAI's unreleased models and systems, which could surface safety issues, influence public trust, and shape future oversight practices.

WHY IT MATTERS

This is a notable expansion of external oversight into OpenAI's unreleased models and systems, which could surface safety issues, influence public trust, and shape future oversight practices.

SOURCES & TIMELINE

1