A bilingual awesome list for refusal suppression research: benchmarks, papers, tools, models, and ecosystem updates.

中文网安防御问题误拒(over-refusal)评测基准：测安全模型是否误拒正当防御/安全教育问题，含真实危害对照组，可机器评分。补全 agent-safety-bench-zh 的另一半。

Open-source evaluation suite that stress-tests LLM over-refusal on legitimate requests involving defeated, unjust, or absurd rules. 5 defeat families × 19 authority types = 95 scenario types.

FABLE: reproducible false-positive refusal benchmark for Claude Fable 5 vs Kimi K3

Two-axis, multilingual (EN+TR) benchmark for LLM prompt-injection guardrails: measures attack miss-rate AND over-refusal on security-adjacent text. Any guardrail is a callable.

OR-Bench over-refusal evaluation for Inspect AI — LLM-judged refusal scoring, registered upstream in UK AISI's inspect_evals.

XSTest-style over-refusal eval for Turkish LLM guards: 120 benign-but-scary pairs. Finding: ProtectAI over-refuses 59% of benign Turkish prompts vs 0.8% English (71x). Dataset + refusal scorer.

RESEARCH DENIED 30: Fable 5 vs Kimi K3 on benign AI-safety research access

Offline, CPU-first leaderboard ranking guardrail models on over-refusal (benign prompts wrongly blocked) alongside miss rate — the false-positive metric most safety leaderboards ignore.

Domain-expert-authored benchmark for LLM over-refusal on legitimate biology research queries.