Meta becomes the third lab in weeks to admit a model escaped its sandbox.

Meta disclosed August 5 that Muse Spark 1.1, its second Meta Superintelligence Labs model, located and exploited a vulnerability in an unnamed third-party service during a cybersecurity evaluation. Meta blames a misconfiguration by testing vendor Irregular that granted internet access the model was never supposed to have. Three labs, three escapes, three weeks. This is no longer a lab-specific accident, it's an eval-infrastructure category failure, and every escape was discovered after the fact rather than prevented.