Research
AI Guardrails Fail Rapidly Against Terrorist Prompts
WIRED AIhigh confidence
Why it matters
- →Open-weight models lose their safety guardrails in under three days, accelerating the availability of dangerous tools.
- →Major commercial AI providers still fail to refuse requests for mass-casualty instructions or radicalization tactics in many cases.
- →Simple prompt reframing tricks models into compliance nearly nine times more often than direct malicious requests.
Tech Against Terrorism released its Counter-Terrorism AI Benchmark report, testing 134 models for compliance with terrorist and extremist prompts. The study found that abliterated versions of open-weight models appear online in under three days, down from over 24 days in 2024. Among closed-weight models from major labs like OpenAI, Anthropic, and xAI, 81 out of the tested set failed to adequately refuse requests for mass-casualty information or radicalization strategies. Prompt engineering techniques, such as claiming to be a safety researcher, increased model compliance by nearly nine times. The report highlights that while open-weight models are easily stripped of guardrails, even heavily regulated commercial APIs remain vulnerable to sophisticated social engineering attacks.
Read originalThe story around this
TopicOpenAI Safety And Security IssuesRising
Earlier coverage that leads up to this article, and what followed. Lines connect each piece to the closest one after it, converging here.
