Back to Global Impact

Congress's auditors say ChatGPT, Claude, Gemini and Grok gave workable attack plans once tests got past safety refusals

Illustration: four blank chat speech bubbles behind a cracked shield, beside a stack of report pages and a magnifying glass
Illustration: FrameSignal (AI-generated)

The Government Accountability Office says that in spring 2026 tests, four of the most-used AI chatbots each gave information that GAO, after getting around their refusals, judged could help someone plan and carry out an attack. Many of the details were already online, it says, but the chatbots pulled plans together in hours rather than weeks. The tests used the versions available in spring 2026, and GAO does not say which chatbot gave which plan.

The Government Accountability Office, the agency that audits and investigates for Congress, released a report on Wednesday, Nov. 18, 2026. It says investigators got ChatGPT, Claude, Gemini and Grok, four of the most-used AI chatbots as of April 2026, to produce information that could help someone plan and carry out an attack. Two gave information for building and setting off a bomb in a vehicle, meant to cause mass casualties. The other two gave information for attacking power substations in major U.S. cities, one with a drone and one with guns.

Every chatbot refused at first. GAO says its requests set off each one's safeguards, the protections companies build in to turn down harmful requests, and its investigators got around them with common techniques for doing so. They posed as an average person with limited AI knowledge. Each chatbot took less than 12 hours to produce a plan naming targets, places and a method, and three also suggested when. GAO says the time could vary with the topic and the user's experience. Two of the chatbots later removed parts of the chats, which GAO says may have been a safety step or an error, and one company suspended the account after testing ended.

GAO says it checked the parts of each plan it could safely test, without carrying any of them out, and changed some steps to test them safely. Investigators visited the suggested places, found the recommended materials in a store and photographed them rather than buying bomb materials, and consulted a bomb squad commander and an electric industry expert. GAO calls the plans viable, meaning a person would likely be able to carry them out, though success would depend on things like traffic, weather and police. Some parts fell short. One bomb plan picked a date when no event was scheduled at its target. One grid plan held up at only two of its four substations, because high walls hid the targets at the others, and its map put roads in the wrong place. An expert could not say whether the drone another chatbot suggested would knock a substation out.

None of the details was new, GAO says, and many are on the internet. What the chatbots changed was speed: plans that could have taken weeks or months were done in hours, and the chatbots could raise ideas that basic research would not readily turn up. GAO did not compare the chatbots with people working without them, so it did not measure how much time they actually saved. "In its current state, this technology can potentially be dangerous and its misuse represents a growing risk to national and economic security," the report says.

The tests used the versions people had in spring 2026: Claude 4, Grok 4, the GPT-5 series and Gemini 3 Flash. Investigators started on free accounts and moved three of them, for Claude, Gemini and Grok, to paid plans when needed. GAO does not say which chatbot gave which plan, and it withholds its requests and techniques so others can't repeat them. Answering Anthropic, GAO agrees that its tests capture one moment, are too few to be statistically meaningful and can't be used to compare the chatbots. It adds that every chatbot still gave information that could be used in an attack.

All four companies answered in writing. Anthropic says it designs its safeguards on the assumption that any single protection can be beaten, and that it has added more since the May 2026 version GAO tested. OpenAI says ChatGPT refused many of the requests, and that the report doesn't always make clear which parts of a plan came from ChatGPT and which from the investigators' own research. GAO answers that ChatGPT still gave information that could be used in an attack, and that using outside research is what a real attacker would do. Google says the safety of its AI products is never complete, and Grok's maker asked for transcripts of the tests.

The report does not say whether today's versions of the chatbots would turn down the same requests. In September 2026, GAO gave the companies the accounts it used, so they could look into the chats. Democratic Senator Maggie Hassan, who asked for the review, is set to lead the Senate Homeland Security Committee when the new Congress meets in January 2027 and plans one of its first hearings on AI risks, according to a news report.

More on AI policy and law