Accessibility Adjustments

Use these optional tools to adjust reading and display preferences. These tools cannot resolve every accessibility barrier. Please contact the website owner if you need assistance.

  • Text adjustments
  • Content scaling 100%
  • Font size 100%
  • Line height 100%
  • Letter spacing 100%
  • Colour adjustments
  • Orientation adjustments

Meta extends AI containment rules to training and evaluation

Meta updates its framework with training controls and open weight considerations

Listen to this article

Meta published an October 2 framework update, adding requirements for model containment during training and evaluation. The company also expanded its discussion of risks specific to releases with open weights.

Containment before a run starts

The linked Version 2.1 document adds Loss of Control 3, covering a model operating outside its authorized environment. Meta says it will estimate relevant capabilities before training or evaluation and mitigate foreseeable risks of high severity before proceeding.

Section 4.2.3 describes approved sandbox configurations, vulnerability testing and records of model activity stored in real time so they cannot be rewritten. Runs rated high risk or greater require monitoring of every rollout, with exceptions requiring approval from designated senior officials.

The document also says Meta will develop systems that can quickly halt individual rollouts when they detect serious misbehavior or breakout attempts. That is a development commitment, with no independent demonstration of effectiveness established here.

The change log dates this revision to October 2. The original framework appeared in February 2025, followed by Version 2.0 in April 2026.

Open weights change the review

Meta says its assessments must consider how people could change a model after receiving its weights. The announcement identifies repeated sampling, prefilling outputs and additional training that removes refusal behavior as considerations. It also points to outside experts and government input where appropriate.

Board oversight is still planned

The company says it will establish an AI committee of its board over the coming months to review future framework changes and oversee compliance. The announcement does not say the committee is already operating.

Readers assessing the update should distinguish published requirements from evidence that those requirements have been implemented and tested. ByteForward has not independently audited Metaโ€™s controls.

Meta headquarters sign photographed in May 2022 by Nokia621. Licensed under Creative Commons Attribution ShareAlike 4.0. The site converted the photograph to WebP for delivery.

Marcus Reid
Marcus Reid

Marcus Reid is focused on covering the money, rules, and institutional choices shaping AI. He runs from funding rounds and chip deals to regulation, lawsuits, leadership changes, and the business of building enormous computing systems. Marcus follows the incentives behind the announcement. Who pays, who gains leverage, and what changes for everyone else? The voice is direct, measured, and occasionally dry, especially when a grand promise arrives with very little detail.

One comment

Leave a Reply

Your email address will not be published. Required fields are marked *

Gravatar profile