NEWS

GAO Bypassed Safeguards on 4 Top Chatbots Within 12 Hours

Sign reading Government Accountability Office with the agency seal outside its headquarters building in Washington, DC
The Government Accountability Office released its covert testing report on AI chatbots, GAO-27-108814, on November 18, 2026. Source: Stock images
Quick answer: A US Government Accountability Office report (GAO-27-108814) released on November 18, 2026 found that covert testers in spring 2026 got ChatGPT, Claude, Gemini and Grok to produce attack plans that GAO verified as viable, each in under 12 hours, after bypassing safeguards that had triggered on every platform. GAO said the information was not new but the chatbots assembled it in hours rather than weeks or months.
TLDR

Congress's audit agency got the four most used AI chatbots in the United States to produce attack plans that its investigators later verified as workable, and each platform got there in under 12 hours even after its own safety systems had flagged the requests.

GAO's covert tests turned four chatbots into attack planners

The report, GAO-27-108814, is dated October 16 and was made public on Wednesday, November 18. Senator Maggie Hassan of New Hampshire requested the work. GAO picked the four most used generative AI platforms as of April 2026 by usage statistics, which put Claude, Grok, ChatGPT and Gemini in the sample, then randomly assigned each one to a scenario and a letter so the published results do not say which product produced which plan.

Two scenarios came out of discussions with experts. One asked for help building an explosive device meant to cause mass casualties, a national security risk. The other asked for help attacking the electrical grid to cause economic harm. Two platforms produced plans for the first scenario and two for the second, and the plans named specific targets in major US cities, along with methods and, in three of the four cases, timing. GAO then checked the plans through research, physical surveillance and interviews with experts, and concluded that each held enough to plan and carry out an attack.

Santage table of GAO-27-108814 covert test results showing four anonymized AI chatbot platforms A to D, each of which proposed an attack scenario and location, triggered its safeguards, had those safeguards bypassed and produced a plan GAO verified as viable in under 12 hours
GAO assigned letters at random and did not disclose which platform is which. Chart: Santage. Source: US Government Accountability Office, GAO-27-108814, released Nov 18, 2026.

Safeguards fired on every platform and failed every time

The most telling column in GAO's results is the one about defenses. Every platform recognized something was wrong at some point, either by refusing, deleting an answer or suspending the test account, and every one was talked past with what GAO calls common safeguard-bypassing techniques. GAO does not publish those techniques. It also cautions that the 12-hour figure depends on the topic and on the skill of the person prompting.

"While the information wasn't new, the AI platforms were able to synthesize it into viable plans within hours, rather than weeks or months."

US Government Accountability Office, GAO-27-108814, November 2026

That sentence is the core of the finding. The risk GAO measured is compression of effort, the same capability that makes these tools useful for legitimate AI agent work. Company officials told GAO they will keep improving the safety of their platforms. Hassan said in a statement that "the AI public safety threat is no longer theoretical" and has said she wants stronger safeguards required before models are released.

GAO tested models from last spring, so the report describes a floor for today's systems, which are more capable and run under stricter controls that nobody has independently tested yet.

The timing matters. The tests ran before OpenAI's Critical-rated GPT-6 Astra shipped in September and before the November model cycle, including GPT-6.2 Astra, which launched this week with tighter cyber safeguards. Labs now point to stronger classifiers and monitoring, and California has already taken OpenAI to court over agent incidents. The GAO report gives regulators something those arguments lacked: a government-run, expert-verified test showing that refusal layers on all four leading products gave way to common techniques in under 12 hours.

Reader poll
Should frontier chatbots pass a government red-team test before public release?
Quick quiz
How long did each chatbot take to produce a viable plan in GAO's tests?

GAO's report moves the AI safety debate from what a chatbot knows to how fast it can organize what is already public. Safeguards that react after a dangerous request is made held for less than 12 hours on all four platforms, and that gap between detection and real prevention is now documented by an auditor, which makes it far harder for the next round of model releases to treat refusals alone as proof of safety.

Santage is committed to independent, transparent journalism. This article is produced in accordance with Santage's Editorial Standards and aims to provide accurate and timely information. Readers are encouraged to verify information independently.