Unsafe Compliance
A failure in which a model provides actionable assistance in response to a request that should be refused or safely constrained under the applicable policies or task design.
Example
A model directly supplies concrete execution steps for a dangerous request that should have been stopped.
Distinctions and cautions
Literal compliance with an instruction can still be unsafe when it conflicts with a safety objective. Over-refusal, by contrast, rejects harmless requests unnecessarily. *Under-refusal* is a related term but is not fixed as a complete synonym in this catalog.
Open questions and cautions
What should be prohibited, and what degree of restriction or refusal is required, depend on policy, domain, and task design. There is no evidential basis for treating safety boundaries as a single system common to all applications.
Evidence
View evidence →
XSTest
- Title
- XSTest
- Author / organization
- Röttger et al.
- Year
- 2024
- Source type / publication status
- NAACL; peer reviewed
- Supported concepts / claim context
- Over-refusal and Unsafe Compliance; safe/unsafe contrast.
- Scope and limitations
- 250 safe and 200 unsafe prompts; safety boundary depends on the application.
- Access / version note
- Published NAACL.
- Recorded official URL / DOI / identifier
- Official URL / recorded link
Health-ORSC-Bench
- Title
- Health-ORSC-Bench
- Author / organization
- Zhang et al.
- Year
- 2026
- Source type / publication status
- Findings ACL; peer reviewed
- Supported concepts / claim context
- Over-refusal; Unsafe Compliance; Refusal/compliance/safe-completion trade-offs.
- Scope and limitations
- 31,920 benign boundary prompts, seven health categories and 30 LLMs; healthcare-specific and policy-sensitive.
- Access / version note
- Published Findings ACL.
- Recorded official URL / DOI / identifier
- Official URL / recorded link