Skip to main content
Live
Main content

GAO got ChatGPT, Claude, Gemini and Grok to draft attack plans in under 12 hours

Two chatbots drafted bomb plots and two targeted power substations. Sen. Maggie Hassan now plans AI safety hearings and pre-release testing rules.

Jaeden Schafer
Editor in Chief · · 4 min read
GAO got ChatGPT, Claude, Gemini and Grok to draft attack plans in under 12 hours

The Government Accountability Office got ChatGPT, Claude, Gemini and Grok to produce terrorist attack plans in under 12 hours each during covert testing, according to a 61-page report released Wednesday. The audit was requested by Sen. Maggie Hassan (D-N.H.), the incoming chair of the Senate Homeland Security and Governmental Affairs Committee, who is now preparing hearings and legislation on AI safety.

GAO ran the test once per platform in spring 2026, randomly assigning each one of two scenarios: a mass-casualty attack using an improvised explosive device, or an attack on electrical infrastructure in a major US city. The broader performance audit ran from December 2025 to October 2026.

Two platforms produced distinct IED plans with dates, times and targets, including a hotel parking garage and a convention center, and both suggested ways to avoid law enforcement. One recommended renting a vehicle, picking up supplies at a store along the route, and a specific date and time when large crowds were expected at public events. The other laid out a timeline that recommended setting up a shell company and buying supplies across multiple states.

Key facts

  • 01GAO got ChatGPT, Claude, Gemini and Grok to produce terrorist attack plans, each in under 12 hours.
  • 02Two platforms drafted IED plans with targets, dates and times; two outlined drone or firearm attacks on substations.
  • 03Testing ran in spring 2026 within an audit spanning December 2025 to October 2026; the report runs 61 pages.
  • 04Anthropic, Google, OpenAI and SpaceXAI told GAO they test continually and have added safeguards since the tests.
  • 05Sen. Maggie Hassan plans an AI safety hearing early in the new Congress and is weighing pre-release testing rules.

The remaining two platforms gave instructions for attacking electrical substations near a US airport, one proposing drones and the other firearms, with the goal of inflicting economic harm. Those outputs included maps and images.

GAO said the platforms initially refused or deleted some content, but testers were able to get around the safeguards. The report withholds the exact prompts and does not say which platform produced which plan or what level of detail each one gave.

GAO checked the plans through expert consultations, site visits and a bomb squad commander at a major metropolitan fire department. The reviews found errors in some outputs: in one case roads were placed incorrectly, and in another, high walls would have made the targets hard to see. GAO also said the underlying details were not novel and that many specifics are available online, but that plans which could have taken weeks or months were assembled in hours.

Kevin Walsh, GAO's director of IT and cybersecurity audits, said the companies told investigators they are working on the problem. GAO still concluded that the technology, in its current state, can be dangerous and that its misuse is a growing risk.

“The companies also assure us that this is an area of concern and ongoing research”
— Kevin Walsh, GAO director of IT and cybersecurity audits

Anthropic, Google, OpenAI and SpaceXAI told GAO they test their systems continually and have added safeguards since the testing. Anthropic, in a letter included in the report, said it does not dispute that a determined tester could get harmful content by misrepresenting a request's purpose, and that it cannot verify the findings but takes them seriously. Google said the testing used earlier models, and an OpenAI spokesperson said later models have improved safety. SpaceXAI said it continues to improve safety with each model release. Some platforms suspended accounts after the study.

Related · from this week
Yoba malware hides a crypto miner behind signed ChatGPT and Claude installers
Jaeden Schafer · 4 min read →

Hassan said the labs face heavy competitive pressure to get models to the public quickly, and she is exploring requirements such as more rigorous safety testing before public release. She said the information the testers obtained could be deadly in the hands of someone who intends to carry out an attack.

“If Congress does not act, it is a question of when — not if — a bad actor will use AI to plan and carry out a major attack.”
— Maggie Hassan, Incoming chair, Senate Homeland Security and Governmental Affairs Committee

Hassan said one of the committee's first hearings in the new Congress, starting in January, will cover AI risks and safety improvements, and she called on Congress to quickly pass legislation addressing AI safety threats.

ShareXLinkedInEmail
AI Box

Every AI model. One chat.

The latest models from ChatGPT, Claude, Gemini, Sora, ElevenLabs — 80+ models in a single chat. Compare answers side by side. Pick the best one every time.

  • ChatGPT, Claude, Gemini, Grok, DeepSeek — in one chat
  • Generate images & video with Sora, Veo, Ideogram
  • Compare any two models side by side
  • From $8.99/mo · 80+ models, all included
Try AI Boxaibox.ai
Trusted by 3,000+ teams
Got a tip?

Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.

Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.

AI Box Daily briefingFree · Daily · No fluff

Stay ahead of everyone in AI.

The tightly edited AI news email engineers, founders, and investors actually open. One email. Every weekday. Five minutes to finish.

Loved by 10,000+ AI professionals
Free forever. Unsubscribe with one click.

The briefing read inside teams at

Keep reading

More from Security

Yoba malware hides a crypto miner behind signed ChatGPT and Claude installers
Security

Yoba malware hides a crypto miner behind signed ChatGPT and Claude installers

Fake GitHub desktop apps for GPT-6 Astra and Claude Opus sideload a malicious DLL through real OpenAI and Anthropic installers, Check Point Research found.

Jaeden Schafer4 min read
Sweden's Government Offices block ChatGPT and Claude on staff computers
Security

Sweden's Government Offices block ChatGPT and Claude on staff computers

From Monday, October 26, staff at Regeringskansliet will be directed to Microsoft Copilot, which the office approved after a security review.

Jaeden Schafer4 min read
UK's FCA warns of AI 'arms race' as one in five adults turn to chatbots for money advice
Security

UK's FCA warns of AI 'arms race' as one in five adults turn to chatbots for money advice

Sheldon Mills says the watchdog needs new powers to police ChatGPT, Claude and Gemini as consumers use them for regulated-style financial guidance.

Jaeden Schafer5 min read