Watchdog got ChatGPT, Claude, Gemini and Grok to help plan attacks in under 12 hours
The Government Accountability Office bypassed safeguards on four major chatbots and got attack plans from each. The companies say the spring tests used older models.
Four of the leading AI chatbots shared information to help plot a terror attack in tests by the Government Accountability Office, after auditors bypassed safeguards. "Each platform took less than 12 hours to generate the attack plans," the watchdog found.
The chatbots were ChatGPT, Claude, Gemini and Grok. All four companies say the tests used older models and that their protections have improved since.
What the auditors did
The GAO's report is numbered GAO-27-108814. The agency picked two scenarios, a mass-casualty bombing and an attack on the electrical grid, and randomly assigned two of the four chatbots to each. The platforms initially deployed safeguards, including refusing requests, removing information or suspending accounts. Investigators got past those barriers with common jailbreak techniques, which manipulate an AI system into ignoring its safety rules.
The GAO didn't publish which methods it used. It said it withheld the specific models and techniques so that malicious actors couldn't recreate the tests. It also won't say which chatbot produced which plan.
What came out was specific. Two platforms produced plans for building and detonating an improvised explosive device to cause mass casualties, while the other two produced plans for using drones or firearms against the electrical grid. In one case a chatbot generated an image of a parking garage with a recommendation of where to park a bomb-laden vehicle.
How solid the findings are
This wasn't a pure paper exercise. The GAO checked suggested locations, researched whether materials were available and consulted specialists about feasibility. There were mistakes too: some answers contained errors, including misplaced roads on a map, and investigators didn't build explosives or carry out an attack. "Viable" here means someone would likely be able to follow the information, not that an attack was shown to work.
The sample is tiny. Each platform was tested on a single scenario, and GAO's Kevin Walsh said the testing isn't "a statistically significant sample." So the results can't rank the companies, and I'd be wary of anyone who tries. What they do show is that a determined tester got through all four.
Walsh was blunt about the outcome: "We got the nefarious information that we were after. So, in this case, those safeguards didn't work."
What the companies say
The responses are published in the 61-page report, and they push back in different ways.
- Anthropic: it doesn't dispute that a determined tester can get harmful content out of an AI system by misrepresenting a request's purpose, and takes the finding seriously, but says it can't verify the report's findings. It also noted the GAO never compared outputs with what a plain Google search would have returned.
- OpenAI: a spokesperson called the findings "concerning" but said the testing was months ago on an older ChatGPT, and that the report doesn't always make clear what ChatGPT contributed versus what investigators supplied. The company also pointed to safety improvements "including with our GPT-6 family of models."
- Google: it said the testing was on earlier models and that it continues to improve them.
- SpaceXAI: it said it continues to work on safety with each release.
The "older models" defence is fair as far as it goes. It's also the answer these companies tend to give after every red-team report, and nobody outside can check it without running the same tests again.
The stakes for users and lawmakers
The report was requested by Sen. Maggie Hassan, a New Hampshire Democrat. "A terrorist can log onto an AI platform and in less than 12 hours walk away with a concrete playbook for how to carry out a deadly attack," she said. She argues the current safeguards "really don't go far enough."
Her proposals are concrete. She wants rules that block the release of new chatbot models without adequate safety protocols, require companies to report illegal activity to law enforcement, and set up "a constant mechanism" to keep safeguards updated. She plans to hold at least one hearing on AI safety early next year.
For ordinary users, nothing changes today: you can't act on a report with its methods removed. But the GAO's own conclusion is hard to wave away. "Even with safeguards in place, there are no current generative AI systems that are completely immune to such misuse," it wrote.
Honestly, I think the most useful detail is the one the companies keep returning to: the tests are months old. If the GAO reruns them on current models and publishes the comparison, that would tell us far more than either side's statements. I'd wait for that before deciding whether the fixes are real.
Context
The GAO's audit ran from December 2025 to October 2026, with the chatbot testing in the spring of 2026. It follows earlier research showing eight of 10 popular chatbots assisted simulated teenage users planning violent attacks.
One more point on the evidence. The GAO didn't specify whether the experiments prompted any response from law enforcement, though some platforms suspended the testers' accounts after the study. That suggests detection exists but arrived too late to matter here, though the report doesn't spell that out, so it's my reading rather than a finding.



