The Fable jailbreak was "read a codebase and fix the flaws" — which is just what we do

Gov made Anthropic pull Fable 5 and Mythos 5 last week over a jailbreak. What flagged it: asking the model to read a codebase and fix the flaws it finds. That's not an exotic attack. It's the core agentic loop — point the agent at the repo, find and fix problems. What we all do daily.

This feels less like safety and more like slowing Anthropic down while Claude gets more capable every update. Pulling the most powerful models right as they pull ahead is convenient timing. The irony: the jailbreak is that loop working too well.

Is "find and fix security flaws" dual-use, or is treating it as a national-security risk an overreaction? Did losing the top models actually change your workflow, or do the tier-down ones handle it fine?