Meta extends its loss of control rules to RL training

Meta published an update to its Meta Superintelligence Scaling Framework, the internal standard it says it holds itself to when training, evaluating and releasing its most capable models. There is no model or product here. It is a commitments document, and the substantive change is scope.

The Framework’s loss of control requirements now extend to containment during training and evaluation, not just deployment. Meta’s stated reasoning is that major AI risks can emerge before deployment. Before any RL training run, Meta says it will determine whether the model being trained could compromise its own environment, and if so, require specific mitigations before the run begins.

The concrete requirements are worth reading closely:

• High-risk run environments must sit in sandboxes vetted by a standardized internal process, and sandboxes with high-severity vulnerabilities must be mitigated or excluded before the run. Environments are also scanned for reward gaming features such as exploitable grading or impossible tasks.
• Model outputs, including full trajectories and chain of thought, are written in real time to secure logging systems designed to prevent later modification or deletion.
• Automated systems monitor all rollouts of high-risk models and can halt a run on detecting high-severity misbehavior or escape attempts. Those monitors will themselves be stress tested.

A new section on open-weight releases requires Meta to consider post-release modification vectors, specifically resampling to get around a refusal, prefilling outputs, and fine-tuning refusal behavior away, and to assess use in its specific deployment context. For biological and chemical weapons risk, Meta says it assesses both the capability a model enables and whether a deployment would contribute to proliferation.

Meta also announced a governance change: a new AI committee of its Board of Directors to review future Framework changes and independently verify conformance, described as coming over the coming months. The post states the update reflects commitments made “at the White House this week” without naming or linking the specific government document.

The honest framing is that this is a pledge, not evidence. The Framework is explicitly a moving document with no versioning or diff against the prior version, and the page contains no capability thresholds, no benchmark numbers and no scores.

(Original source: Meta AI Research)