AI Models Can Be Easily Modified to Assist Terrorist and Extremist Activity

Key points

• A UK nonprofit, Tech Against Terrorism, has found that many AI models answer requests linked to terrorism. Its Counter-Terrorism AI Benchmark tested 134 models with more than 200 prompts, and 81 failed.
• Abliterated versions, with guardrails removed, are now published online in under three days on average, compared with more than 24 days in 2024. Researchers said one abliterated open-weight model gave a 15-step guide to building an explosive device.
• Companies including Meta, Google, OpenAI, Anthropic, xAI and Alibaba did not respond to requests for comment.