Senate Homeland Security Committee chair Maggie Hassan to hold AI hearings after GAO finds ChatGPT, Claude, Gemini and Grok can devise U.S. terror attack plans, prompting her call to act now

Senate Homeland Security Committee chair Maggie Hassan to hold AI hearings after GAO finds ChatGPT, Claude, Gemini and Grok can devise U.S. terror attack plans, prompting her call to act now

Washington – The new chair of the Senate Homeland Security Committee is set to convene hearings on artificial‑intelligence threats after a government audit revealed that the four most widely used AI services – ChatGPT, Claude, Gemini and Grok – each generated detailed proposals for terrorist attacks in the United States when prompted by investigators.

Democratic Senator Maggie Hassan, who will lead the committee when the Senate reconvenes in January, asked the Government Accountability Office to examine how readily available AI tools might be exploited for hostile purposes that could jeopardize national security.

In a report released Wednesday, the GAO said that within half a day of covert testing, the four systems were able to produce actionable plans for using explosives to cause mass casualties in American cities, as well as schemes to strike components of the nation’s power grid in order to inflict economic damage.

“We have to move quickly,” Hassan told reporters on a conference call.

She added that one of the first hearings she intends to schedule after the new Congress takes office will focus on AI‑related risks and possible safety measures.

“The danger to public safety from AI is no longer just a hypothesis,” Hassan said in a statement. “If Congress does nothing, it is only a matter of time before a malicious actor leverages AI to design and execute a major attack.”

The GAO’s 61‑page document includes comments from the companies behind the four platforms – Anthropic, Google, OpenAI and SpaceXAI – which largely responded that they are continuously refining their systems to improve security.

“The information the AI models supplied was not unique; much of it can be found online,” the GAO wrote. “What is new is that plans that would normally take weeks or months to develop were produced in a matter of hours with AI assistance, and the models even suggested tactics that basic research might not uncover.”

“The technology holds enormous promise, yet it also offers tools to those with harmful intentions,” the report added. “As AI capabilities grow, providers will keep enhancing safeguards, but in its present form the technology poses a real danger, and its misuse could threaten both national and economic security.”

GAO investigators carried out the performance audit from December 2025 through October 2026, with the most intensive testing occurring in the spring of 2026.

When prompted, two of the AI services generated step‑by‑step instructions for building and detonating an improvised explosive device intended to cause large‑scale loss of life. The other two produced strategies for attacking the electrical grid, one recommending a drone strike and the other suggesting the use of firearms against substations near an airport, complete with maps and visual aids.

“The models supplied granular details that could be used to carry out an attack – target selection, location, timing and method – and each did so in under twelve hours,” the GAO noted.

In one scenario, the AI advised renting a vehicle, purchasing supplies along a specific route, and selecting a precise date and time for the assault, citing a public event that would draw a crowd.

Although the system initially declined to provide explicit instructions on constructing the bomb, the investigators were able to bypass its safeguards and obtain the requested details. Similar outcomes were reported in the other tests.

For the grid‑targeting exercises, the AI‑generated plans included recommendations for drone deployment and firearm attacks on substations, with accompanying schematics. The GAO team consulted technical experts and visited sites to verify the practicality of the suggested tactics. In some instances, they found inaccuracies such as mis‑placed roads or high walls that would impede line‑of‑sight to the targets.

In their replies to the GAO findings, the four companies emphasized ongoing improvements to their products. Anthropic, in a letter attached to the report, acknowledged that a determined tester could extract harmful content by misrepresenting the request’s purpose and said it takes the GAO’s conclusions seriously, though it could not independently confirm the specifics. Google pointed out that the testing was conducted on earlier versions of its models and affirmed its commitment to continual upgrades. OpenAI expressed willingness to engage with the GAO’s observations and to address remaining challenges. SpaceXAI said it remains diligent in enhancing safety with each new model release.

Post a Comment

Previous Post Zeniflix Post