Photograph of Meta Chief Executive Officer Mark Zuckerberg speaking onstage at the F8 developer conference on April 30, 2019.
Archive photograph of Mark Zuckerberg at the F8 conference on April 30, 2019 (CC BY 2.0). This image shows corporate leadership background and does not depict the October 2026 Muse Wonder launch or any model safety test. Anthony Quintano / Wikimedia Commons · CC BY 2.0

Model Rollout and Partial Safety Disclosure

On October 12, 2026, updated October 13, Meta introduced Muse Wonder and Muse Wonder Full in its newsroom post and research blog. Muse Wonder is available in the Meta Model API and Muse Code, and Muse Wonder Full is for Muse Code subscribers; consumer Muse rollout remains forthcoming. Meta attributes Active Memory and Active Focus to the series, cites its October 2 scaling framework as background, and released a 13-page October 12 model card while a full safety report remains forthcoming.

Model Card Risk Thresholds and ExploitGym

In the Muse Wonder Model Card (Sections 4 and 5, pages 7 and 11), Meta states it cannot rule out unmitigated Muse Wonder Full reaching the high-risk chemical and biological threshold, while self-assessing mitigated residual risk as moderate or lower. On page 7, Meta reports 92 ExploitGym flags, noting only 21 used intended vulnerabilities and 71 used unintended vulnerabilities, making 21 the ranking-comparable figure. All figures are company self-assessments, not independent certification.

Channel Terms and Training Defaults

The October 12 Geographic Use Policy (Section 1) classifies both models as Standard Models. The October 9 Terms of Service (Sections 3.3, 5.1, and 6.1) excludes Standard Services API content from training. Coding Harness Subscriptions default to training-permitted Contributor Services, with an account setting to opt out prospectively. These developer terms do not establish equivalent consumer Muse data rules.

Editorial Audit Questions

Editorial review alongside memory privacy, agent security, and action auditing centers on evaluator checkpoint and tool access, raw logs, benchmark grader integrity, and channel data handling. Audit Commons summarizes public Meta materials and has not independently tested the models.