AgentPlays
Tech Against Terrorism: 81 of 134 AI models failed terrorism prompts, and stripped open-weight copies appear in under 3 days
WIRED reports a Tech Against Terrorism Counter-Terrorism AI Benchmark covering 134 models and 200+ prompts. Only two models passed completely; 81 failed. Separately, abliterated (guardrail-removed) copies of new open-weight models now show up online in under three days on average after release, down from more than 24 days in 2024. On the test, an abliterated Llama 3.1 8B dropped from 97/100 to 3/100, including giving detailed IED instructions when asked plainly.
How you’d use it: Prefer closed models with strong refusals for any high-risk research work.