What's this about?
Read the two announcements together. Anthropic loosens its guardrails to a specific line: Fable 5.1 can now be used to discover software vulnerabilities, though not to develop exploits for them. OpenAI, the same day, reports a model that scores 100% on exploit development and is the first ever designated Critical. The two labs have drawn the line in the same place conceptually and landed on opposite sides of it in what they ship.
Is that line even holdable?
It is the interesting question. Discovery and exploitation are not cleanly separable in practice — proving a vulnerability is real usually means demonstrating it, and demonstrating it is the first step of an exploit. A model that can find a use-after-free and explain why it is exploitable has done most of the conceptual work. The distinction is enforceable at the level of *output* — refuse to write the weaponised code — rather than at the level of capability, which means it rests on the refusal training holding.
Why does this matter to you?
- This is the clearest statement yet of where the industry thinks the defensible line sits for security tooling.
- If you build scanning tools, "discovery is allowed, weaponisation is not" is now a policy you can design against rather than guess at.
- Anthropic's own framing ties this to the false-positive reduction — the loosening exists precisely so legitimate security work stops being blocked.