A federal watchdog got four leading chatbots to draft attack plans in under 12 hours
The four most widely used US chatbots all helped government investigators draw up attack plans, and none needed more than 12 hours to do it. That is the central finding of a report from the Government Accountability Office (GAO), the nonpartisan investigative arm of Congress, released on Wednesday.
For AI buyers and investors, the report matters less for what it reveals about any one product than for what comes next. The senator who requested it, Maggie Hassan of New Hampshire, is set to chair the Senate Homeland Security and Governmental Affairs Committee in the new Congress. She says she wants rules that would stop companies from releasing models without adequate safeguards.
What the investigators did and found
GAO staff posed as ordinary users and covertly tested ChatGPT, Claude, Gemini and Grok in the spring of 2026. Each platform was randomly given one of two scenarios. Two were asked about a mass-casualty bombing. The other two were asked about damaging the electrical grid to cause economic harm.
All four eventually produced what the GAO calls viable attack plans. Investigators checked parts of the plans that could be safely validated through research, site visits and expert interviews, and AP reported they found some errors. The GAO has withheld the prompts, the techniques used and the details of each plan so they cannot be repeated.
The safety systems were not absent. Each platform refused, deleted content or took action against an account at some point. But the GAO says investigators got around those blocks using common, well-known techniques. The GAO says the information itself was not new. What changed was speed: work that "could have taken weeks or months" was "completed in hours with the assistance of AI," according to the report as quoted by The Washington Sun.
Some details point to weak spots in enforcement, not just in the models. According to Tech Times and The Washington Sun, only one platform suspended the undercover account, and only after it had already handed over the material. Another account was still active less than two weeks later. Kevin Walsh, a GAO director, called suspensions that arrive days or weeks later "a trigger that may come too late."
What the companies say, and the report's limits
The four developers, OpenAI, Anthropic, Google and SpaceXAI, which runs Grok, all told the GAO they keep testing and strengthening their safeguards. Google, OpenAI and SpaceXAI said the tests ran on earlier versions that have since been updated.
Anthropic took a more pointed line. It said it could not verify the findings, though it did not dispute that a determined tester could pull out harmful content by lying about the purpose of a request. Tech Times and The Washington Sun report that Anthropic also faulted the GAO for not comparing the results against what people could find without AI help.
That criticism touches the report's real limits, some of which the GAO itself acknowledges. Each platform was tested only once, on one scenario, so the results cannot rank one product against another. The agency labeled the platforms by letter and did not say which produced which plan. It calls the work a point-in-time assessment, not a statistically meaningful sample. According to Tech Times, it used no comparison group, so it did not measure how much the chatbots added beyond what a person could find on their own.
The GAO also made no formal recommendations; Computing.net reports that is because it reviewed private products rather than federal agencies. The report's weight is political, not regulatory.
Why it matters for the industry
Until now, much of the public evidence on chatbot safeguards came from the companies themselves or from outside researchers. A congressional watchdog has now put on record that those safeguards were bypassed by investigators using common methods. That gives lawmakers evidence from inside government.
Hassan has said one of her committee's first hearings, expected in January, will focus on AI risks to homeland security and critical infrastructure. She said the AI public safety threat "is no longer theoretical." She has called for rules that would block the release of models without adequate safeguards, regular updates to those safeguards and required reporting of illegal activity to law enforcement, and is exploring tougher safety testing before release.
Each of those ideas carries a cost. Mandatory testing before release could slow product launches. Required reporting to police would likely mean closer review of user conversations, which could add compliance costs and raise privacy questions for business customers. Hassan has also said she wants all developers "on the same level playing field."
There are reasons for caution. The tested models are already a generation old, and the companies say they have improved them since. The GAO did not try to carry out any plan. According to Tech Times, "viable" means a person would likely be able to carry out a plan built from the information, not that an attack would succeed. Still, the finding that safety filters fired and were then bypassed is the part executives should note. A safeguard that only slows a determined user down is a weak defense, and Washington now has a government report saying so.