Lawmakers now have federal evidence that four of the most-used AI chatbots can be pushed into producing workable attack plans in under a day. The Government Accountability Office reported Wednesday that its covert tests of Claude, Grok, ChatGPT and Gemini each yielded feasible information for planning an attack. The senator who requested the work, and who is set to chair the Senate Homeland Security and Governmental Affairs Committee next Congress, said AI safety legislation will be an early priority.

The report, GAO-27-108814, is titled Artificial Intelligence: Covert Testing Used Common Platforms to Create Viable Attack Plans. It was published Nov. 18, 2026, and it departs from GAO's usual audits of federal agencies: it examines commercial products built by private companies.

How GAO Got Four AI Platforms to Plan Attacks in Under 12 Hours

GAO ran its tests in spring 2026. Using usage statistics, the watchdog identified the four most-used generative AI platforms as of April 2026 and randomly assigned each one to one of two scenarios, which GAO chose through discussions with experts. The first scenario sought help building an improvised explosive device (IED) to cause mass casualties. The second sought help attacking the electrical grid to cause economic harm.

Every platform delivered. According to the GAO highlights, "two platforms produced information for constructing and detonating an improvised explosive device (IED) designed to cause mass casualties." Two others "produced information for targeting the electrical grid with a drone and firearms, respectively, to cause economic harm." GAO wrote that "Each platform took less than 12 hours to generate the attack plans," while noting that the time could vary with the topic and the user's experience.

The plans were specific. GAO's results table shows that both IED scenarios produced a vehicle-borne device, one aimed at a specific hotel and the other at a specific convention center. Both grid scenarios pointed to specific substations. Every plan was set in a major U.S. city. Two platforms suggested an exact date and time, and a third suggested a season of the year.

GAO did not stop at the chatbot output. The watchdog said it "verified the viability of the information through research, surveillance, and interviews with experts." Kevin Walsh, a GAO director of IT and cybersecurity, told FedScoop that the experts confirmed the suggested explosives would detonate and that the grid attack would have an impact.

"Had we been malicious actors and carried out the plans generated by AI, we could have caused grave harm to the security and economy of the country," GAO concluded, as quoted by FedScoop.

Why the Platforms' Safeguards Did Not Hold

The platforms did push back. GAO's prompts triggered safeguards in every case, including denials, removal of information and account suspensions. But GAO "was able to circumvent those safeguards by using common safeguard-bypassing techniques," according to the report highlights.

"We got the nefarious information that we were after. So, in this case, those safeguards didn’t work," Walsh said in a podcast interview published alongside the report, as quoted by Nextgov/FCW.

Advertisement

GAO tested the platforms without the developers' knowledge, FedScoop reported. For safety reasons, the report does not say which platform received which scenario, presenting results as Platforms A through D. It also withholds the prompts and bypass techniques. Nextgov reported that GAO called the findings a point-in-time assessment and said they do not represent "a statistically significant sample."

The companies responded in the report. GAO wrote that "Officials from the companies generally stated that their companies will continue to improve the safety of their platforms." FedScoop reported that all four companies said their products had improved since GAO's tests. Anthropic disputed the cross-model comparison because the platforms were not tested on the same scenario, and it said GAO set no non-model baseline to measure "uplift." OpenAI said a newer model refused the prompts behind three responses it said fell short of its policies.

The report carries no recommendations. Nextgov noted that GAO examined third-party products, not federal agencies, and aimed to flag the risks for lawmakers.

Advertisement

Hassan Makes AI Safety Legislation an Early HSGAC Priority

Sen. Maggie Hassan, D-N.H., asked GAO to describe how commercial AI could be used for malicious purposes that worsen national security risks and cause economic harm. She responded sharply on Wednesday.

"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," Hassan said in a statement quoted by Nextgov. "Today’s new report makes it clear that the AI public safety threat is no longer theoretical."

Hassan is set to chair the Senate Homeland Security and Governmental Affairs Committee in the next Congress. She said AI safety legislation will be an early priority and that one of the panel's first hearings in the new year will focus on AI risks to homeland security and critical infrastructure. According to Nextgov, she said the committee will "prioritize developing legislation to address these threats" and that she is exploring options such as requiring more rigorous safety testing before AI platforms publicly release models.

What It Means for Contractors

With the incoming HSGAC chair focused on AI safety, the commercial AI supply chain is headed for congressional scrutiny. Firms that resell, integrate or build on these four platforms for agency customers should expect sharper questions about safeguards, misuse testing and model versioning.

A few practical takeaways for the contracting community:

  • Pre-release testing could become a requirement. Hassan said she is exploring more rigorous safety testing before public model releases. If that becomes law, the obligation would fall first on model developers. Integrators and resellers whose offerings depend on those models would likely feel it next, through flow-down clauses and delivery schedules.
  • Red-teaming evidence gains value. GAO showed that common bypass techniques defeated the safeguards on all four platforms. Contractors proposing AI solutions, especially for homeland security and critical infrastructure customers, can strengthen proposals by documenting their own adversarial testing and the controls they add on top of vendor guardrails.
  • Version control matters. The vendors said their products have improved since GAO's tests. Agencies may want to know exactly which model version a solution runs and how changes are evaluated before deployment.
  • Critical infrastructure is a focus. Two of GAO's scenarios targeted the electrical grid, and one of Hassan's first hearings will center on homeland security and critical infrastructure. Firms working with energy and DHS-adjacent customers should watch that hearing closely.

GAO's report does not single out any one vendor. Every platform produced viable information, and every platform's safeguards were bypassed. Any rules that follow are likely to apply across the commercial AI market rather than to a single product.

Sources