NewA calmer, faster ZeroHour — security assessments on articles, story timelines and search from any page.

#harmbench

0 stories this week · 5 this month · 5 in 6 months · mostly AI

Stories per day · 30 days

Relatedai-safety2 jailbreak2 abliteration1 activation-steering1 advbench1 adversarial-prompts1 alignment1 benchmarks1 black-box1 evals1 guardrails1 harmful-refusal1 helm1 jailbreakbench1 llm-refusal1 llm-safety1 llm-security1 over-refusal1

5 stories, newest first · one card per story

arXiv cs.AI / cs.LG / cs.CL9d ago

WRDO learns weighted directions to change LLM refusal behavior

The method is tested on three open-weight models with HarmBench and can also be reversed to restore refusal.

llm-refusal alignment

2

Hugging Face daily papers13d ago

Guardrail safety scores track policy labels more than detection capability

Ten arms trained on about 53k dual-labeled prompts show HarmBench-copyright scores spanning 0-100 while detection stays unchanged.

guardrails safety-benchmarks

1

arXiv cs.CR14d ago

ECHO jailbreak search scores reply harmfulness instead of compliance prefixes

Authors report up to 0.898 StrongREJECT and 89.23% HarmBench success on five open-weight models; code is public.

jailbreak llm-safety

1

arXiv cs.CR16d ago

Diagnose Before You Refine: Multi-Dimensional Judge Feedback for Query-Efficient Jailbreak Attacks

A new paper reports a query-efficient jailbreak method with 89.8% HarmBench success in 10 queries.

jailbreak llm-security

2

arXiv cs.AI / cs.LG / cs.CL22d ago

Searching for "Harmful Refusal": A Psychometric Audit of an AI Safety Benchmark

A psychometric audit finds HarmBench and HELM Safety do not isolate a single harmful-refusal attribute.

2