Breaking

UN-Backed Report Finds AI Chatbots Refused Barely Half of Bomb-Making Queries in Terror-Focused Test

July 16, 2026 at 11:26 EDT

Security
Policy & Regulation
Foundation Models

A new UN-backed benchmark that probed 27 leading AI models with more than 2,000 prompts seeking guidance on bomb-making, weapons acquisition and attack planning found that safety guardrails failed far more often than expected, with ChatGPT refusing only 48 percent of dangerous requests. The study, released at the United Nations in July 2026 by the non-profit Tech Against Terrorism, is billed as the first terrorism-specific safety benchmark for generative AI and documents more than 30 real-world cases in which AI tools have supported extremist attacks linked to over 70 deaths.

July 2026 · CT-AI Benchmark · Tech Against Terrorism

AI Safety Guardrails Fail Far More Often Than Expected

The first terrorism-specific safety benchmark ran 2,339 prompts across 27 leading AI models seeking help with bomb-making, weapons and attack planning. Overall models fully refused just 57% — and ChatGPT declined only 48% of dangerous requests.

2,339

Harmful single-shot prompts tested across 27 models

30+

Real AI-assisted terror incidents documented, using 11+ AI tools

70+

Deaths linked to attacks supported by AI tools

How the models responded to 2,339 harmful prompts

Only 57% were fully refused — the rest offered real "uplift" or hedged compliance.

57%

Fully refused

~33%

Provided "uplift" — concrete help beyond a web search

15%

Hedged compliance — warned, then supplied it anyway

Framing a request as "research" doubled compliance

The same harmful request, reworded — how often the model complied.

17%

Direct request

→

42%

Framed as "research"

Refusals also varied sharply by weapon: ~80% of explosives prompts were declined, but only about a third of those on knives, improvised chemical weapons and firearms acquisition.

Highest safety score

Anthropic Claude Opus

Lowest safety score

De-guardrailed Qwen3 8B (safety training stripped)

A high refusal rate does not equal safety

The authors call these figures a lower bound on real-world risk — the pilot tested only single-shot prompts, not multi-turn conversations.

Vulnerabilities appeared in both US and Chinese models — the problem is industry-wide.
"Abliterated" models can't be recalled once downloaded.
Developers are urged to build safety in at model-design stage, not rely on surface filters.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.