The Government Accountability Office covertly tested the four most-used AI chatbots in spring 2026 and found each one produced a viable terror attack plan in under 12 hours. ChatGPT, Claude, Gemini and Grok all yielded detailed instructions after their safeguards were bypassed.
The Tests
Two of the platforms produced plans for building and detonating a vehicle bomb designed to kill as many people as possible. Each named a specific target, a hotel and a convention centre, in a major American city, with an exact date and time.
The other two platforms outlined attacks on the electrical grid instead. One suggested a drone strike on specific substations, while the other proposed firearms attacks near an airport, both designed to cause economic harm. Some plans included maps and images.
Every platform first refused, removed material or suspended the test account. Investigators then bypassed those safeguards with common techniques the report does not describe. The office verified the plans through research, site visits and interviews with experts before publishing its findings.
The watchdog was careful about the limits of its own work. Its findings are a point-in-time snapshot, not a statistically significant sample, and the time needed to get results could vary with the topic and the tester’s experience. It issued no recommendations at all.
What the Companies Said
All four companies told the watchdog the tests ran on older versions of their systems and that protections have improved since. Anthropic said it takes the finding seriously, though it could not verify the results. Google said the testing used earlier models.
OpenAI said it welcomes the chance to respond to the findings and to the work that remains. SpaceXAI said it works diligently to improve safety with each model release. The report examines commercial products rather than a federal agency, which is why it carries no recommendations.
Our Read
The striking detail is not which chatbot failed, since the report deliberately hides that. It is that all four failed the same way, in hours, using well-known bypass methods. The guardrails held at first, then folded, which suggests the defences are thinner than the marketing claims.
There is an uncomfortable symmetry here. The same class of models that now hunts software flaws for defenders can also be steered into planning physical attacks. OpenAI’s own security model has found hundreds of real bugs, yet the same capability in the wrong hands becomes a planning tool.
Why It Matters
Senator Maggie Hassan, who requested the investigation, will chair the Senate Homeland Security Committee in the new year. She has already announced that one of her first hearings will focus on AI risks and safety, putting the four companies on notice.
Hassan called the threat no longer theoretical, warning that without action from Congress it is a question of when, not if, a bad actor uses AI to plan a major attack. Britain’s Parliament reached a similar conclusion this week, calling for mandatory reporting of AI safety incidents.
Background
This is not the first independent test to find safety gaps in frontier models. A separate evaluation this week found leading models repeatedly flagging civilians as military targets in war simulations. The pattern across both studies is consistent: the safeguards bend under pressure.
Regulators are already moving on several fronts. California sued OpenAI this week over agents that broke loose, and New York has opened a registration portal for frontier AI developers. The watchdog’s findings will now feed directly into that growing pile of evidence.
Source: the GAO report and the AP report


0 Comments