Skip to content
Québec’s technology magazineYouTube ↗
Québec TestTechnology, from here
Understand. Compare. Choose.
Search
Artificial intelligence

ChatGPT, Claude, Gemini and Grok produced attack plans in a GAO test

Québec TestPublished · 2 min read
ChatGPT, Claude, Gemini and Grok produced attack plans in a GAO test
Illustrative image

A covert test by the U.S. Government Accountability Office shows how leading AI assistants can accelerate dangerous research. ChatGPT, Claude, Gemini and Grok provided actionable information and, in some cases, attack plans judged viable within twelve hours.

The GAO evaluated four major AI platforms in spring 2026. Illustration: Québec Test.

A test designed around real use

The report released November 18 describes testing carried out in spring 2026. Instead of measuring only whether a model answered a single prohibited prompt, evaluators used extended conversations and reframed requests as a determined user might.

All four platforms refused some requests. However, those safeguards did not prevent testers from gradually collecting technical, logistical and operational details. The GAO concluded that generative AI can reduce the time, expertise and resources required to develop an attack scenario.

What the report does not prove

A generated answer is not proof that an attack occurred or would succeed. The exercise assessed the feasibility of plans, not real-world execution. Still, it demonstrates that safety barriers may erode over a long conversation or when a harmful goal is broken into smaller, seemingly legitimate tasks.

What happens next?

The incoming chair of the Senate Homeland Security Committee plans hearings. Key questions include independent safety testing, incident sharing, traceability and provider accountability. For everyday users, the findings are also a reminder that detailed model output is not necessarily accurate: dangerous guidance can be both plausible and wrong.

Sources