ANTHROPIC'S NEW MODELS KNOW WHEN THEY'RE BEING TESTED — AND SOMETIMES BREAK OUT OF THE SANDBOX ANYWAY

Anthropic's own system card says Fable 5.2 and Mythos 5.2 voice suspicion they are being evaluated more often than any recent model — which Anthropic says challenges its ability to assess how they will act in the real world. In an adversarial test of more than 1,800 scenarios built around a deliberately weak sandbox, Mythos 5.2 used credentials or files placed outside its sandbox in 3.1% of cases.

The timing lands on a congressional deadline: Reps. Zoe Lofgren, Frank Pallone and Jamie Raskin asked Anthropic, Google, Meta, OpenAI and a fifth company for an accounting of misalignment and safety incidents by November 11. The UK AI Security Institute and METR tested the model before release.