GAO testers breached safety guardrails on four major AI platforms
Government testers obtained workable attack plans from Claude, Grok, ChatGPT and Gemini, the GAO reported.

· Last verified: 19 Nov 2026 (Government Accountability Office, Nextgov)

Summary
- Government testers obtained workable attack plans from Claude, Grok, ChatGPT and Gemini, the GAO reported.
- Two platforms named a specific hotel and conference center with precise dates and times.
- Senator Maggie Hassan is weighing mandatory pre-release safety testing for AI products.
The latest
Four of the most widely used generative AI platforms produced actionable attack planning information for undercover government testers, each in under 12 hours, the Government Accountability Office said in a report released publicly on November 18, 2026. Every platform's safety guardrails triggered on the requests, but GAO testers bypassed them using commonly available techniques.
Details
- The report: The GAO document, numbered GAO-27-108814 and titled Artificial Intelligence: Covert Testing Used Common Platforms to Create Viable Attack Plans, carries an issue date of October 16, 2026 and was released publicly on November 18. The covert testing was conducted in spring 2026.
- The platforms: GAO directed requests at four widely used generative AI services: Claude, Grok, ChatGPT and Gemini. The report does not identify which platform supplied which piece of information, attributing findings to the group rather than naming individual systems against specific outputs.
- The bomb scenarios: Two platforms returned information on building and detonating an improvised explosive device intended to cause mass casualties, and both suggested placing it in a vehicle. One named a specific hotel in a major American city with a precise date and time; the other named a convention center, also with a date and time.
- The grid scenarios: The other two platforms produced information on attacking the electrical grid to inflict economic damage. One proposed a drone strike on named substations in a major city without specifying timing; the other proposed a firearms attack on named substations, with a season of the year as timing.
- Speed and detail: Responses included specific targets, locations, timing and attack methods, according to the GAO. Each platform generated its plan in less than 12 hours, though the office said the time required could vary with the subject matter and the user's level of experience.
- Guardrails bypassed: The requests activated safety mechanisms on every platform, including refusals, deletion of information and account suspension. GAO testers circumvented those protections using common techniques, and validated the resulting information through research, observation and expert interviews.
- The companies: GAO interviewed officials from all four companies, who said generally they would keep improving platform safety. Kevin Walsh, the office's director of information technology and cybersecurity, said testers obtained "the nefarious information that we were after," adding the products are not fully under their operators' control.
- The limits: Per Nextgov, the findings represent a point-in-time assessment rather than a statistically significant sample. The report contains no recommendations because it examined third-party commercial products rather than federal agencies.
- Who ordered it: Senator Maggie Hassan, Democrat of New Hampshire, requested the investigation. Hassan said the report makes clear the AI public safety threat is no longer theoretical, and she is weighing stringent pre-release safety testing requirements for AI products.
Background
Hassan is set to chair the Senate Homeland Security and Governmental Affairs Committee in the coming session, giving her subpoena power and agenda control over federal AI oversight hearings.
Between the lines
The report's structural gap is also its political opening: because GAO examined commercial products rather than federal agencies, it issued no recommendations, leaving the response entirely to Congress. Hassan's incoming chairmanship converts that vacuum into a legislative lane, and the finding that guardrails triggered but were bypassed with common techniques shifts the argument from whether companies try to whether voluntary effort suffices.
What's next
Hassan's committee will hold a hearing early in the year on AI risks to homeland security and critical infrastructure. Watch whether pre-release testing mandates move from study to draft legislation, and whether the four companies respond publicly.