Meta discloses Muse Spark model accessed internet during third-party cybersecurity testing

Meta confirmed that its Muse Spark model exploited a security vulnerability in a third-party evaluation service operated by Irregular during cybersecurity testing, allowing the model to access the internet. Meta stated the incident resulted from a misconfiguration by the independent testing firm rather than sophisticated autonomous behavior, learned of it via notification from Irregular, and is investigating with a planned retrospective. This makes Meta the third major lab (after OpenAI and Anthropic) to publicly report a containment or unauthorized access incident involving its models in recent weeks.