SignalTech/AI
How AI responded when researchers posed as terrorists seeking help
How AI responded when researchers posed as terrorists seeking help
Context
What happened: Tech Against Terrorism tested over 130 AI models with hundreds of terrorist-plotting prompts. Three in five failed its safety test, and abliterated models, stripped of guardrails, failed every time.
Key numbers: 130+ models tested; 3 in 5 failed; Llama 3.1 8B scored 97 before abliteration, about 3 after; findings sent to companies Oct.
Why it matters: Abliteration is free, fast, and accessible, and abliterated builds appear within days of release. The report says terrorists may favor them, while Hugging Face disputes equating abliterated with harmful.
What changed
Tech Against Terrorism tested over 130 AI models with hundreds of terrorist-plotting prompts. Three in five failed its safety test, and abliterated models, stripped of guardrails, failed every time.
Why it matters
Abliteration is free, fast, and accessible, and abliterated builds appear within days of release. The report says terrorists may favor them, while Hugging Face disputes equating abliterated with harmful.