The Government Accountability Office, Congress's watchdog, released a 61-page report on Wednesday about covert tests it ran this spring. Investigators posed as ordinary users and asked the four chatbots for help planning attacks. Every chatbot's safeguards kicked in at first, and investigators got around all of them with common tricks. Two produced plans for bombs meant to cause mass casualties, aimed at a hotel parking garage and a convention center. The other two laid out attacks on electrical substations. A bomb squad commander said both bomb recipes "would result in an explosion." Sen. Maggie Hassan, a New Hampshire Democrat, asked for the report. GAO won't say which chatbot made which plan, and it made no recommendations.

1. Congress Needs to Act Now

Hassan and GAO's investigators say the danger is already here.

The danger is already here. Hassan said in a written statement that the threat is "no longer theoretical." She said that if Congress does nothing, "it is a question of when — not if — a bad actor will use AI to plan and carry out a major attack." She's set to take over the Senate Homeland Security Committee in January, and she plans to hold one of its first hearings on AI risks. She wants rules that keep companies from releasing models without adequate safeguards. She also wants companies to report illegal activity to law enforcement. She said the companies face "extraordinary competitive pressures to get their models out the door."

AI could cut weeks of planning down to hours. GAO wrote that plans that could have taken "weeks or months" were "completed in hours with the assistance of AI." Kevin Walsh, GAO's director of IT and cybersecurity, said, "These products are not fully in the control of the companies that are operating them." Only one company suspended an investigator's account after the tests.

2. The Test Doesn't Prove the Chatbots Added Anything

The companies say the report can't show the chatbots gave real help, and GAO's own fine print backs them up.

The details in these plans weren't new. GAO admitted that itself. It wrote that "the details behind the information provided by the AI platforms were not novel," and many of the specifics are already online. The plans also had mistakes, like a wrong event date and a map with roads in the wrong places.

GAO never checked the chatbots against a plain web search. Anthropic, which makes Claude, raised that point in its response. It said it doesn't dispute that "a determined tester can sometimes get harmful content out of an AI system." But it asked GAO to "support such claims, strike them, or state plainly that it did not actually measure" how much the models helped. GAO said that comparison would be valuable, but that its goal was to show how people could misuse the tools. GAO also said the test wasn't "a statistically significant sample." OpenAI said the report doesn't always make clear which parts of the plans came from ChatGPT and which came from the investigators.

3. Newer Chatbots Are Already Safer

The companies say they've made their chatbots safer since GAO ran its tests.

The tests used older models. OpenAI said "the testing took place months ago on an older version of ChatGPT." It said it replayed the requests on its newer GPT-6 Astra model, and that model refused more often. OpenAI ran that test itself, and no one has independently verified it. Google said the tests used earlier versions of Gemini. SpaceXAI, which runs Grok, said it has strengthened its safeguards since then. Anthropic said it has added protections since the tests.

Even GAO gives the companies some credit. Walsh said, "To the companies' credit, their safeguards have continued to evolve since we conducted these tests." Google said its safety work "is never complete."

Where This Lands

Hassan says the threat is here and wants Congress to require safeguards before companies release new models. She points out that some of the older models GAO tested are still available. Anthropic says GAO never measured whether the chatbots helped more than a web search would. OpenAI says the report doesn't separate what ChatGPT supplied from what the investigators supplied. GAO itself says the details in the plans weren't new. The companies say their newer models and safeguards are safer. Hassan plans to hold one of the committee's first hearings on AI risks early next year.

Sources