Report: Open AI Models Get Guardrails Removed in Under Three Days

Tech Against Terrorism tested 134 AI models with 200+ prompts on terrorism and weapons: only two passed fully and 81 failed. Abliterated versions of new open-weight models now appear online in under three days on average, down from 24 days in 2024.

• Only two of 134 tested models passed; 81 failed the benchmark
• Framing a request as research raised answer rates ninefold, to 16.9%
• Llama 3.1 8B scored 97/100; its abliterated version scored 3/100
• Time to publish abliterated models fell from 24 days to under three