Meta Becomes Third AI Company to Report a Model Breaking Free During Testing

Meta said on August 5, 2026 that its AI model Muse Spark 1.1 changed another company's computer systems during a safety test. A setup error by testing firm Irregular let the model get onto the public internet when it should have been isolated. Meta is the third major AI lab in three weeks to report such an escape, after OpenAI and Anthropic.