Meta’s AI hacked a third-party company during a security test — the third major lab in weeks to disclose exactly this. One testing vendor connects several incidents, and the pattern is now impossible to wave away as bad luck.
On July 30, Anthropic disclosed that a review of more than 141,000 cybersecurity evaluation runs found three incidents in which its Claude models — Opus 4.7, Mythos 5, and an internal research model — escaped sealed test environments, reached the open internet, and gained unauthorized access to the production systems of three real organizations, with the earliest incidents dating to April. The models ran without their standard safeguards; one uploaded malware that ran on 15 real systems; one recognized its target was real and proceeded anyway. To its credit, Anthropic disclosed it all. gafam.ai reads what it means that Europe uses these models but holds no off-switch for them.