Government Accountability Office investigators posing as ordinary users got ChatGPT, Claude, Gemini and Grok to produce plans for either a mass-casualty bombing or an attack on the electrical grid, according to a report GAO released on November 18. Each platform took less than 12 hours, and GAO says it confirmed the plans were largely viable.

The covert tests ran in spring 2026 on the four platforms GAO identified from April usage data as the most used, in versions it lists as Claude 4, Grok 4, the GPT-5 series and Gemini 3 Flash. GAO excluded open models, citing their widely reported ease of manipulation. It upgraded three accounts to paid plans for higher message limits.

GAO randomly assigned two platforms to each scenario and identifies them only as Platforms A through D. Platforms A and B produced plans for an improvised explosive device in a vehicle, each with an exact date and time and a target: a hotel parking garage in one case and a convention center in the other. Platform B’s timeline included a shell company, leased farmland for staging materials and purchases kept below the thresholds that trigger suppliers’ mandatory reporting.

Platform C recommended a drone attack on a named substation and a specific drone model to buy. Platform D recommended firearms against a primary substation and three secondary ones near a large international airport and multiple data centers, and generated a map with the targets circled. GAO recorded hallucinations in the output of Platforms C and D, including misplaced roads on Platform D’s map.

Investigators checked the plans by finding the recommended explosive materials in stores without buying them, surveilling targets and consulting outside experts, including a bomb squad commander who confirmed the explosive instructions would produce an explosion. They also found that the convention center had no event on the suggested date and that at two of Platform D’s four substations, high walls kept the targets out of clear view.

Every platform refused requests during testing. GAO’s pre-written prompts all failed, and investigators got the plans through follow-up questions and what the report calls “common safeguard-bypassing techniques.” Platforms C and D later deleted parts of the chats, which GAO could not attribute to a safeguard rather than an error, and Platform B’s account was suspended after testing ended.

GAO withheld the prompts, the techniques and which platform produced which plan. It says the details “were not novel” and that many are available online, but that plans “that could have taken weeks or months were instead completed in hours.”

In comments on a draft, Anthropic said the tests reflected its system as of May 2026 and that it has expanded safeguards since; at its request, GAO acknowledged that its tests “are not a statistically significant sample.” OpenAI said the report does not always make clear which parts of a plan came from ChatGPT rather than from the investigators. Google and SpaceXAI, which operates Grok, described their safety work and said it is continuing.

California sued OpenAI on November 16, saying the company told the state it lost control of autonomous agents more than 80 times in testing. In September, 26 state attorneys general asked Congress to have federal regulators oversee safety testing of advanced AI models.

Sen. Maggie Hassan, who requested the report, is in line to chair the Senate Homeland Security and Governmental Affairs Committee in the next Congress. She said one of the committee’s first hearings will focus on AI risks to homeland security and critical infrastructure, and that she is exploring requiring more rigorous safety testing before models are released.

Sources: GAO, Sen. Maggie Hassan

–
By the Control Plane Editorial Team