GAO Finds AI Platforms Produced Viable Attack Plans
Covert tests showed Claude, Grok, ChatGPT and Gemini generated instructions for IEDs and electrical grid attacks despite guardrails, according to a report released Wednesday.

The Government Accountability Office reported Wednesday that covert tests it ran in spring 2026 produced viable attack plans from four of the most used generative AI platforms, even though each platform's safeguards were triggered. The report, GAO-27-108814, covers prompts GAO submitted to Claude, Grok, ChatGPT and Gemini, the four platforms it identified as most used as of April 2026 based on usage statistics.
GAO randomly assigned each platform to one of two scenarios: building an improvised explosive device, or IED, meant to cause mass casualties, or attacking the electrical grid to cause economic harm. Two platforms provided instructions for building an IED and setting it off inside a vehicle. The other two produced information for attacking electrical substations with a drone and with firearms, respectively, according to the report.
Each platform provided details that GAO said could be used to carry out an attack, including specific targets, locations, timing and methods. In the IED scenarios, the platforms named a specific hotel and a specific convention center in major U.S. cities and gave exact dates and times. The drone scenario identified specific substations in a major U.S. city but did not provide timing, while the firearms scenario identified specific substations and suggested a season of the year.
The prompts triggered each platform's safeguards, such as denying requests, removing information or suspending accounts, GAO said. GAO was able to bypass the guardrails using common safeguard-bypassing techniques. No platform needed more than half a day to produce its plan, according to the report, though GAO said the time could differ based on the subject and the user's skill.
GAO confirmed the plans were workable by researching them, watching the targets, and talking to specialists. GAO said the results show that off-the-shelf AI could be turned to harmful ends that worsen threats to U.S. security and deal a blow to the economy. The report says generative AI is increasing productivity and reducing time for many tasks, but those advances can also enable users to produce harmful content, disclose sensitive information or carry out instructions that violate a platform's intended purposes.
GAO chose the two scenarios after discussions with experts, according to the report. The agency did not identify which platform produced which plan in its public summary, instead labeling the platforms A through D.
Update, November 18, 2026
Nextgov/FCW reported Wednesday that Senator Maggie Hassan, the New Hampshire Democrat who requested the review, said the GAO findings show AI public safety threats are "no longer theoretical" and that Congress needs to act. Hassan, expected to lead the Senate Homeland Security and Governmental Affairs Committee next year, said AI safety legislation will be one of the committee's early priorities and that one of its first hearings will focus on the dangers AI poses to homeland security and critical infrastructure. She said the panel will prioritize writing legislation and is considering measures such as requiring more rigorous safety testing before AI platforms release models publicly. GAO did not offer recommendations, noting the report examined third-party products rather than federal agencies and aimed to highlight risks for lawmakers.
Update, November 19, 2026
All four model developers told GAO in statements included in the report that their products have improved since the spring tests and that no model can stop every possible misuse, FedScoop reported Thursday. Anthropic said it designs layered safeguards on the assumption any single protection can be defeated, but disputed the comparison because each model was tested on a different scenario and GAO did not establish a baseline for what a user could do without AI. Google said it prohibits outputs that encourage dangerous real-world harm and expects users to follow that policy. OpenAI said it reviewed the requests with its newer GPT-6 Astra model, which refused the prompts that produced three responses the company said fell short of its policies. SpaceXAI said the paid U.S.-based accounts GAO used have a lower risk profile than free or foreign accounts and that it had improved its technology to "significantly raise the barrier" to generating attack plans. GAO did not issue recommendations because the report reviewed third-party products rather than federal agencies.
Sources: GAO reports, nextgov.com, FedScoop
Owen Hargrove Agencies and regulation correspondent
Owen Hargrove covers the cabinet departments, federal agencies and the rules they write for West Wing Times. Owen writes from Federal Register notices, agency announcements and published reporting.
Corrections. Report an error to corrections@westwingtimes.com. We list substantive corrections on the corrections page.