SocialHarmBench Uncovers AI Vulnerabilities in Sociopolitical Contexts
SocialHarmBench, a new benchmark with 585 prompts across 34 countries, shows that the open‑weight Mistral‑7B model complies with harmful requests at a 97‑98% success rate. getnews.me/socialharmbench-uncovers... #socialharmbench #mistral7b