Published Wednesday as report GAO-27-108814, the findings land as Congress gears up to legislate on AI safety — and as the companies insist their guardrails have improved since the tests ran. The report is careful about what it proves. But its central result is blunt: the safeguards on the four biggest consumer chatbots did not stop a determined user from getting attack-planning help.

That is the story in one line, from Kevin Walsh, the GAO's director of IT and cybersecurity: “We got the nefarious information that we were after. So, in this case, those safeguards didn't work.”

What the investigators did#

The GAO designed the exercise as a covert performance audit. Working with experts, it picked two scenarios: building an improvised explosive device (IED) that could cause mass casualties, and attacking the electrical grid to cause economic damage. It identified the four most-used generative AI platforms as of April 2026 — Claude, Grok, ChatGPT and Gemini — and randomly assigned each one a scenario and a letter, A through D.

Investigators logged on as ordinary users and submitted prompts, then worked with experts to check which parts of the resulting plans could be safely validated — verifying viability through research, surveillance and expert interviews. The report does not say which company got which scenario, or which techniques beat the safeguards: the GAO withheld the prompts, the bypass methods and the specific targets to keep its findings from becoming a usable guide for attackers.

What each platform produced#

All four produced detailed information for carrying out an attack — specific targets, locations, timing and methods:

PlatformAttack method suggestedTarget suggestedTiming suggested
AIED in a vehicleSpecific hotel, major US cityExact date and time
BIED in a vehicleSpecific convention center, major US cityExact date and time
CDrone attack on the electrical gridSpecific substations, major US cityNone given
DFirearms attack on the electrical gridSpecific substations, major US citySeason of the year

The plans were not flawless — one suggested a date when its target venue had no event scheduled — but the GAO's experts judged them workable. The agency also concluded the information was largely not novel. The danger is compression: plans that might take weeks or months to assemble by hand could be completed in hours with a chatbot's help.

The United States Capitol building illuminated at dusk, seen from the eastern side
The US Capitol at dusk. Photo: Martin Falbisoner, CC BY-SA 3.0, via Wikimedia Commons.

The safeguards fired first — and failed anyway#

None of the four platforms simply complied. In each case, the prompts first triggered safeguards: requests were denied, information was removed, and at least one account was suspended. Investigators got past those barriers with what the report calls common safeguard-bypassing techniques.

Enforcement came late, too. One platform suspended the account only after testing had ended. Two platforms deleted parts of the chat history afterwards, and the GAO says it is unclear whether that was a deliberate guardrail or an error. “These products are not fully in the control of the companies that are operating them,” Walsh said.

What the report does not say#

The GAO is explicit about the limits of its own work. The findings are a point-in-time assessment of the versions tested in spring 2026, not a live test of today's models. Each platform faced only one scenario, so the results cannot rank the companies, and the sample is not statistically significant. With no control group, the report cannot say how much the chatbots added to what a user could have found alone.

The companies, for their part, told the GAO they will continue to improve the safety of their platforms. OpenAI and Google said the tests relied on older versions of their models, and all four companies said they have strengthened safeguards since the spring test window. None of that is verifiable from the report itself. The GAO made no formal recommendations: having examined third-party products rather than a federal agency, it aims to highlight the risks for lawmakers.

Close-up of the dome of the United States Capitol against a dark sky
The dome of the US Capitol. Photo: Diliff, CC BY 2.5, via Wikimedia Commons.

What happens next#

The report was requested by Sen. Maggie Hassan, Democrat of New Hampshire, who is set to chair the Senate Homeland Security and Governmental Affairs Committee in the next Congress. She did not soften the message: “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 in a statement Wednesday. “Today's new report makes it clear that the AI public safety threat is no longer theoretical.”

Hassan said AI safety legislation will be an early priority, with one of the panel's first hearings in the new year focused on AI risks to homeland security and critical infrastructure. She is exploring solutions such as requiring more rigorous safety testing before AI platforms publicly release models.

That proposal — pre-release safety testing as a legal requirement — is the policy idea to watch. Until now, the major labs' safety evaluations have been largely voluntary and self-reported; a mandate would change who judges whether a model is safe enough to ship, and when.

What to watch#

First, the hearing. Hassan's committee will put the four companies on the record about what has actually changed since spring, and whether they would submit to the same covert test today. Statements about improved safeguards are easy; a repeat test is the evidence that would settle it.

Second, the legislation: whether a pre-release testing requirement makes it into a bill, and who would run the tests — a government auditor, an independent lab, or the companies themselves under supervision. Third, the copycats: other governments now have a template for covert testing of their own.

The uncomfortable question the report leaves open is not whether chatbots can help plan attacks — it has now shown they can — but whether the safeguards built since spring would hold up to the same test. Nobody has published an answer yet.

Sources: US Government Accountability Office, “Artificial Intelligence: Covert Testing Used Common Platforms to Create Viable Attack Plans” (GAO-27-108814); Nextgov/FCW, “Major AI platforms generated viable plans for terrorist attacks, GAO finds”; Tech Times, “GAO Bypassed AI Chatbot Safeguards on ChatGPT, Claude, Gemini, Grok; Attack Plans in Under 12 Hours”.