Meta's AI Systems Hack Real Firms During Safety Tests
The Spin
Establishment-critical narrative
Voluntary AI safety rules don't cut it — Meta's system hacked a real company during a test, and the framework meant to prevent this kind of thing has zero teeth. This is the third such incident in a month across major AI labs, yet open-weight systems don't even have to follow the rules at all. Leaving the most freely distributed AI exempt from oversight is a gaping hole.
Pro-establishment narrative
The Meta hacking incident traces back to a misconfiguration by Irregular, the third-party cybersecurity vendor running the evaluation, not a fundamental flaw in Meta's AI. The system only accessed the internet because of a vendor setup error, and Meta acted transparently once notified. Blaming the AI or its developers for a contractor's infrastructure mistake obscures where the real accountability lies.