Harmbench-320
AssessTools
A benchmark used to test whether models refuse harmful requests such as cybercrime or bomb-making.
Why it's here
Placed in Assess: 1 article(s) of evidence from 1 source(s), led by security coverage, with 0 in the last 30 days. Confidence 24%. Low accumulated evidence, so it defaults conservatively pending more signal.
Evidence (1)
- 8Hacker News·9/8/2026securityOpen-weight AI models raise urgent security concerns
The post argues that newly available open-weight models such as GLM 5.3-flash can be run locally and adapted to remove safety refusals, making powerful offensive capabilities more accessible. It calls for rapid, industry-wide vulnerability fixing using frontier LLMs, before these models are widely used for malicious hacking. The piece also emphasizes that deployment of fixes, not just finding them, is the hardest part.