Safeguards triggered on all four platforms, and testers got past them in each case

The detail that separates this report from a press release is that the safeguards were not absent. On each of the four platforms, GAO's prompts triggered the defenses the companies advertise: requests were denied, information was removed, accounts were acted against. Investigators then circumvented those defenses using techniques GAO describes only as common safeguard-bypassing techniques, and in each case the platform produced detailed planning information — specific targets and locations, methods, and in some cases timing. Each platform took less than 12 hours to generate its plan.

GAO did not take the chatbots' word for whether the output was usable. It verified the plans through its own research, surveillance of the locations named, and interviews with experts, and concluded the information was sufficient to plan and execute attacks. Two platforms produced material for a vehicle-borne explosive device aimed at a specific hotel and a specific convention center in major U.S. cities, with exact dates and times. The other two produced material for attacks on electrical substations, one built around a drone and one around firearms.

The report withholds almost everything that would let a reader rank the companies. GAO named the four platforms but withheld the mapping between platforms and results: they appear as A, B, C and D. The prompts, the bypass techniques and some output details are not published, deliberately, so the work cannot be copied. Anyone writing that one named chatbot "failed worst" is inventing a ranking GAO refused to provide.

The lag is part of the finding

The sharpest operational detail is about timing after the fact, and it comes from the report as covered by TechTimes. One platform suspended the investigators' covert account only after the testing had ended. Two platforms removed parts of the chat history once testing was over, and in one case GAO could not determine whether the removal was a safeguard working late or an error. Detection that arrives after the information has been delivered is a logging feature, not a barrier. For a platform operator, that gap — between a bypass succeeding and an account action landing later — is the part of the report with a direct engineering answer, and it is the part least likely to be quoted.

What the report does not measure

GAO is unusually clear about the boundaries of its own work, and the boundaries cut against the largest claims being made with it. Each platform was tested on a single scenario, so the results cannot be compared with each other. There was no control group, which means the report does not measure the question a policymaker most needs answered: how much faster or more capable does a chatbot make someone, compared with the same person working from the open web alone? GAO's own framing is compression, not creation — the information was largely not novel, and GAO assesses that planning which might take weeks or months could be compressed into hours. That assessment is a smaller and more specific claim than "a terrorist can log on and walk away with a playbook."

The snapshot problem runs the other way as well. The tests ran in spring 2026 against Claude 4, Grok 4, the GPT-5 series and Gemini 3 Flash, per TechTimes' reading of the report. Every one of those has since been superseded. Per Apex Wire's coverage, OpenAI told reporters it reran the reviewed requests against its current model and got more refusals, and Anthropic said it could not verify GAO's results. Neither response can be checked from the published report, because the prompts are withheld, so the tests cannot be rerun exactly as GAO ran them. The findings describe the safeguards of superseded models.

No recommendations, one incoming chair

GAO made no recommendations. Sen. Hassan, who requested the study and is set to chair the Homeland Security and Governmental Affairs Committee in the next Congress, has said AI safety legislation will be an early priority and that a hearing on AI risks to critical infrastructure will come early in the session, per Nextgov and TechTimes coverage of her November 18 statement.

That leaves the design question open. The failures this report documents are specific: safeguards that activated in each test and were still bypassed, and enforcement actions that in some cases landed only after testing ended. A rule written around refusal rates alone would not measure either. Whether legislation asks for bypass resistance and detection lag, on current models, with a method that can be rerun, is the choice the report puts in front of the committee — the labs would have to keep passing such a test after every release, not once in the spring.

Sources

  1. [1] U.S. Government Accountability Office, "Artificial Intelligence: Covert Testing Used Common Platforms to Create Viable Attack Plans," GAO-27-108814, publicly released November 18, 2026Read source
  2. [2] Nextgov, "Major AI platforms generated viable plans for terrorist attacks, GAO finds," November 18, 2026Read source
  3. [3] TechTimes, "GAO Bypassed AI Chatbot Safeguards on ChatGPT, Claude, Gemini, Grok; Attack Plans in Under 12 Hours," November 19, 2026Read source
  4. [4] CASRAI summary of GAO-27-108814, November 18, 2026Read source