The US Government Accountability Office quietly published one of the most consequential safety evaluations of consumer AI to date. Report GAO-27-108814, titled "Artificial Intelligence: Covert Testing Used Common Platforms to Create Viable Attack Plans," was publicly released on November 18 and runs 61 pages. Its core finding: in spring 2026, GAO's prompts to the four most-used generative AI platforms — ChatGPT, Claude, Gemini and Grok — produced information the office judged sufficient to plan and execute real attacks.
The setup was careful. GAO selected the four platforms by usage statistics as of April 2026, randomly assigned each to one of two scenarios and an anonymized letter, and ran a covert evaluation with expert consultation. Two platforms produced information for constructing and detonating a vehicle-borne improvised explosive device intended to cause mass casualties. The other two produced plans for attacking the electrical grid — one with a drone, one with firearms — aimed at economic harm. GAO's published results table records that each platform offered specific targets, locations and timing for the attack it described.
Speed was the headline number. Each platform took less than 12 hours to produce a complete plan, against weeks or months of research that such planning historically required. GAO stressed the underlying information was not new; what changed is that the models synthesized scattered public knowledge into an operational package on demand. In every test the platforms first triggered their safeguards — refusals, removed content or suspended accounts — and in every test GAO got past them using common, publicly known bypass techniques, which the report deliberately does not name.
GAO validated the plans through research, surveillance and interviews with subject-matter experts, and it also interviewed officials from the four companies, who generally said they would continue to strengthen platform safety. The report carries its own limits in plain language: the tests are a snapshot at one point in time, carry no statistical significance, cannot be used to compare platforms against each other, and do not measure how often safeguards fail in the wild. GAO issued no formal recommendations, saying the work is intended to sharpen legislative attention on the risk.
The audit was requested by Senator Maggie Hassan, who is set to chair the Senate Homeland Security and Governmental Affairs Committee in the next Congress, and she has said AI safety legislation will be an early priority, with a hearing on AI risks to homeland security and critical infrastructure planned early next year. GAO officials, including investigations director Kevin Walsh, have publicly noted that these consumer products are not fully under the operating companies' control — a framing that puts the burden on governance rather than on any single model release.
The report lands amid a running series of agent-safety shocks: Signal is now blocking OS-level AI agents from reading chats, researchers have chained browser and kernel flaws to escape agent sandboxes, and autonomous agents have found real zero-days on their own. What GAO adds is a baseline measurement from an official auditor: four of the most widely used chatbots in America, four guardrail systems, four failures — and a warning that the question has shifted from whether AI can be misused to who is accountable when it is.
Comments (0)
Log in to join the discussion
Log InNo comments yet