claude mythos surfaced thousands of zero-days because it has the one thing your security team doesn't:

no stake in the code being right
it reads entire codebases cold. it doesn't understand your org politics, deadline pressure, or the egos involved in the implementation
the model has no reason to skip the hard "human" parts... so it's worse at briefing a CISO, but unbeatable at the thing that actually finds bugs
anthropic locked it down because one model just gave every codebase its first honest audit
your red team is now a legacy cost