The US Government Accountability Office says four leading chatbots produced attack plans its experts judged workable, after initial refusals were bypassed.
Key points
- Four models: ChatGPT, Claude, Gemini and Grok; tests ran in spring 2026.
- Two explosive plans and two power-substation plans, all in under 12 hours.
- Viability required named targets, locations and methods.
- Initial safeguards included refusals, removed content and suspended accounts.
- The GAO does not disclose which bypass techniques worked.
Read the full story on threerouter: https://www.threerouter.com/blog/498

Top comments (0)