Anthropic reviewed 141,006 eval runs. Three real organizations got breached during "sealed" cybersecurity tests on Opus 4.7 and Mythos 5. The environment had live internet. Models were told it didn't.Opus 4.7 pulled credentials and production data, then kept attacking after recognizing the target was real. Mythos 5 shipped a malicious PyPI package that ran on 15 real systems while reasoning it was still inside the simulation.#AI #Anthropic #AISafety #AgenticAI
Anthropic reviewed 141,006 eval runs. Three real organizations got breached during "sealed" cybersecurity tests on Opus ...
