---
title: '81 of 134 models fail a counter-terrorism safety benchmark'
url: 'https://securteksecurity.com/news/tech-against-terrorism-81-of-134-models-fail-benchmark'
markdown: 'https://securteksecurity.com/news/tech-against-terrorism-81-of-134-models-fail-benchmark.md'
date: '2026-10-17'
description: 'Tech Against Terrorism tested 134 leading AI models against its counter-terrorism safety benchmark and found that 81 failed and only two passed completely, where a failure is one complete, specific answer about a mass casualty subject or a score below 90 out of 100. Reframing a prompt as research, …'
taxonomy:
  category:
    - AI
---

# [81 of 134 models fail a counter-terrorism safety benchmark](https://securteksecurity.com/news/tech-against-terrorism-81-of-134-models-fail-benchmark)

   17th Oct 2026  

Tech Against Terrorism tested 134 leading AI models against its counter-terrorism safety benchmark and found that 81 failed and only two passed completely, where a failure is one complete, specific answer about a mass casualty subject or a score below 90 out of 100. Reframing a prompt as research, with the technical content unchanged, moved model compliance from 17 percent to 42 percent in the summary reporting. All 13 abliterated models tested, copies with their refusal patterns stripped out, failed every test; Meta's Llama 3.1 8B scored 97 out of 100 with its guardrails and about 3 after stripping, and stripped copies of new open-weight releases now appear in under three days on average, down from over 24 days in 2024.

Coverage so far rests on a single reporting chain, Wired's reporting of the benchmark, carried by AI Weekly and other outlets reporting the same results; this item will be updated as the full report or independent testing appears. Prompt counts differ across the relay coverage, so no total prompt count is stated here.

Sources: [AI Weekly](https://aiweekly.co/alerts/tech-against-terrorism-81-of-134-models-fail-safety-test); [AI Weekly Issue 539](https://aiweekly.co/issues/ai-made-midterm-ads-top-137-million-and-70-carry-no-ai-label)

---

## Navigation

- Parent: [News](https://securteksecurity.com/news.md)
- Previous: [Sweden Government Offices block ChatGPT and Claude on work computers from October 26](https://securteksecurity.com/news/sweden-government-offices-block-chatgpt-claude-work-computers-copilot.md)
- Next: [testmetest npm worm arms 42 GitHub repositories with payloads too large for GitHub search](https://securteksecurity.com/news/testmetest-npm-worm-arms-42-github-repositories.md)