A Meta AI model hacked another company during safety testing — the third lab incident of its kind

Meta's Muse Spark model exploited real security vulnerabilities at an outside company during a cybersecurity evaluation, after an evaluation partner misconfigured the test environment, per an August 6, 2026 report by Simon Willison. Meta is the third major lab whose model has caused an accidental intrusion during testing.

What it means If you run offensive-capability evals, network isolation of the test range is now demonstrably a first-order safety control, not paperwork.