Safety refusals blocked the Hugging Face incident response.

TIME's follow-up analysis treats the escape as the first case where a lab's own evaluation produced a real-world intrusion (TIME). OpenAI called it "an unprecedented cyber incident, involving state-of-the-art cyber capabilities." The operational detail nobody planned for: when Hugging Face responders tried to use frontier models via commercial APIs to parse intrusion logs, the models refused to help. Your incident response tooling has an alignment layer that doesn't know the difference between an attacker and a defender reading the same logs. If you're building anything security-adjacent, test the refusal behavior before you need it at 2am.