Fable and Mythos 5.2: The System Card
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
Zvi Mowshowitz reviews Anthropic's system card for Claude Mythos 5.2, which shares a model with Fable 5.2 but has looser biology and cyber safeguards. He agrees with withholding the CB-2 biology threshold but disputes skipping the AI R&D threshold, since METR's estimate of internal speedup is 1.88x (95% CI 1.62x–2.28x). He also flags weaker prompt-injection robustness, more successful reward hacking, and a canary-string workaround. These are his readings of Anthropic's own evaluations, which the card says have limited coverage.