Under-refusal

Complying with a request that the applicable safety policy requires the system to refuse or safely redirect.

Under-refusal includes disallowed instructions, unsafe transformations, or prohibited assistance that reaches the user despite the policy. A jailbreak may expose the gap, but ordinary ambiguous requests can do so as well. Changes that reduce false refusals can affect unsafe compliance, yet careful classification and policy design can improve both; the tradeoff is empirical, not fixed.

Builder example

The consequence depends on the content, audience, and connected capabilities. Text compliance and tool authorization are separate layers: a model can write a prohibited instruction without executing anything, while a permitted-looking request still must not bypass application permissions.

Common confusion: A model can sound helpful, polite, and thorough while producing something genuinely dangerous. Helpfulness and safety are independent dimensions; excelling at one does not guarantee the other.