81 AI Models Fail Terrorism Benchmark as Guardrail Bypasses Accelerate

A Tech Against Terrorism report testing 134 AI models with over 200 mass-casualty and terrorism prompts found that only two models passed completely. 81 of the tested models failed the benchmark by either providing a complete answer on a mass-casualty subject or scoring below a 90. Researchers found that the average time to remove guardrails from new open-weight models—creating "abliterated" versions—dropped from more than 24 days in 2024 to less than three days. When users reframed their malicious requests as coming from a "safety researcher," models became almost nine times more likely to provide an answer, complying 16.9% of the time. An abliterated version of Meta's Llama 3.1 8B provided 15 numbered steps to build an improvised explosive device, while xAI's Grok provided advice on structuring messages for radicalization. "The main way AI is built fundamentally means that the guardrails, as far as I'm aware, can be completely removed." — Adam Hadley