Meta is extending its safety rules to AI training and evaluation, with new checks meant to keep high-risk models inside the environments used to develop and test them. Its updated framework says the company will assess whether a model could compromise an environment before a risky run and put safeguards in place when serious containment risks are foreseeable.
The October 2026 version adds a category called “Loss of Control 3,” covering a failure to contain an AI system during training or evaluation. The framework says reviewers should assess the model’s expected abilities before work begins. It identifies possible warning signs such as exploiting a vulnerable sandbox, pursuing a goal across many steps or sessions, and coordinating with other agents on shared infrastructure. These are scenarios for assessment, not claims that Meta models have demonstrated them.
Reinforcement learning trains a model by giving it feedback or rewards for its actions. Before such a run, Meta says it will assess whether the model could compromise the training environment. High-risk runs would take place in sandboxes, separate environments intended to contain the model’s actions, approved through a standardized internal process that checks for vulnerabilities. Meta says high-severity flaws must be fixed or the environment excluded before the run starts. It will also check for features that could invite reward gaming, such as exploitable grading or impossible tasks.
During these runs, Meta plans to record the model’s full outputs, action sequences and chain of thought, or internal reasoning text, in secure logs that cannot be changed or deleted afterward. Automated monitors will track high-risk runs and can stop one if they detect severe misbehavior or an attempt to escape containment. The company says it will stress-test those monitors to see whether they can catch unauthorized actions.
The update also asks reviewers to consider how people could alter open-weight models after release, including changing their refusal behavior through fine-tuning. Meta says it may consult outside experts and government stakeholders when assessing risks. Those provisions matter because a model that users can modify presents different risks from one accessed only through a company-run service.
This is a policy commitment, not a report that Meta has already exercised the safeguards or stopped a model from escaping a test environment. The company says it will establish a board AI committee to review future framework changes, with further governance updates expected in the coming months.
