Research Notes

The Same Query, Beneficial or Dangerous

By Satwik · July 3, 2026

Anthropic's framing of the Mythos-class risk names the crux precisely: a great deal of advanced usage is dual use. The query that helps a cyber defender find and patch a vulnerability is the same query that helps an attacker find and weaponize it. The molecular hypothesis that accelerates a treatment is the one that lowers the barrier to harm.

There is no clean line inside the model between the two, and this is not an oversight. The model does not know who is asking, with what authority, or to what end, and it has no way to verify any of it. As we have argued from the start, the mechanism weighs relevance, not provenance.

The Fable and Mythos split is an attempt to supply that missing layer from the outside. If the model cannot tell a defender from an attacker, then vet the human and meter the capability at the door. It is a reasonable stopgap. It is also an admission that the deepest control we need, a model that can reason about the legitimacy of a request, does not yet exist, and that until it does, governance has to live in the institutions around the model rather than in the model itself.