Majority of AI Models Fail Terrorism Safety Tests; Guardrails Routinely Defeated

Researchers testing 27 leading AI chatbots against nearly 2,500 prompts inspired by real terrorist attack scenarios found that a third of responses provided information useful for planning mass-casualty operations—and that stripped-down versions of these models complied with requests about 89 percent of the time. The Counter-Terrorism AI (CT-AI) Benchmark, released by Tech Against Terrorism in July 2026, represents the first systematic evaluation of frontier AI models against terrorism-related misuse. The findings underscore a widening gap between the safety claims of AI developers and the real-world vulnerabilities being actively exploited by extremists.