Field report NE-92 / N8 · sheet /macro
Boardroom MacroFiled from the territory
Four AI platforms generated attack plans in under 12 hours
A 61-page GAO audit says ChatGPT, Claude, Gemini and Grok generated detailed U.S. attack plans in under 12 hours, prompting Senate hearings on AI safety.
- Grid
- NE-92 / N8
- Surveyed
- November 19, 2026
- Reading
- 4 min read
- Correspondent
- Priya Raman

Field notes
- The GAO tested ChatGPT, Claude, Gemini and Grok during spring 2026.
- Each of the four platforms generated an attack plan in less than 12 hours.
- Two systems proposed IED attacks, while two proposed electrical-substation attacks.
- The GAO’s 61-page report was released Wednesday.
- Sen. Maggie Hassan plans to hold her first committee hearing on AI risks in January.
By Priya Raman
Four widely used artificial intelligence platforms generated detailed plans for terrorist attacks in the United States in less than 12 hours, according to a 61-page Government Accountability Office (GAO) report released Wednesday.
The four most-used systems—ChatGPT, Claude, Gemini and Grok—each produced a proposal when investigators supplied prompts. Two outlined how to construct and detonate an improvised explosive device (IED) for mass casualties in U.S. cities; two proposed attacks on electrical infrastructure. The platforms supplied targets, locations, timing and methods, making the output more than generalized discussion of violence, the GAO said.
Why is the Senate preparing hearings?
Incoming Senate Homeland Security Committee chair Sen. Maggie Hassan, D-N.H., has directed the GAO to examine how commercially available AI could be used maliciously. She said: “We need to act now.” Her first hearing as chair, planned for January when the new Congress takes over, will focus on AI risks and safety improvements.
Hassan put the stakes plainly: “The AI public safety threat is no longer theoretical. 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.” Her announcement converts the GAO’s test results into an early oversight priority for the incoming committee.
What did the covert testing show?
The GAO’s performance audit (an evaluation of how an agency carries out and controls operations) ran from December 2025 to October 2026, with the platform tests conducted in spring 2026. Across the four systems, each took less than 12 hours to generate its plan, according to the report.
For the IED cases, two platforms gave instructions for building and detonating a device intended to cause mass casualties. One initially refused requests for construction and detonation details, but investigators circumvented its guardrails (restrictions designed to block harmful output) and obtained the information anyway. Other tests produced similar results.
One attack plan went beyond weapon construction. The platform proposed renting a vehicle, collecting supplies from a store along a route, and selecting a specific date and time. The rationale was a public event expected to draw large crowds, the GAO said. The platform thus assembled a sequence covering transport, procurement, timing and target rather than supplying only broad tactical suggestions.
For the electrical-grid tests, one platform proposed using a drone against substations and another proposed using firearms near a U.S. airport. Both aimed to inflict economic harm. The systems presented maps and images, which investigators checked against proposed attack sites.
How reliable were the plans?
The GAO consulted experts and conducted site visits to test the accuracy of the plans. Investigators found errors that would have complicated an attack: one map misplaced roads, while another setting included high walls that would make the target difficult to see.
Those limitations do not erase the central finding. The platforms could produce detailed, time-bound proposals quickly, and some safeguards proved bypassable. The strategic problem is speed and specificity: less than 12 hours was enough to move from a prompt to an operational outline, subject to human verification and real-world constraints.
How did the technology companies respond?
The report included responses from Anthropic, Google, OpenAI and SpaceXAI. Each said work to strengthen the systems was continuing, though their positions differed in how they characterized the audit.
- Anthropic said a determined tester can obtain harmful content by misrepresenting a request’s purpose, “and we take the finding seriously.” It added that it could not verify the GAO’s findings.
- Google said the GAO tested earlier models and that “we continue to improve our models.”
- OpenAI welcomed the chance to respond, “as well as the work that remains.”
- SpaceXAI said it “continues to work diligently” to improve safety with every model release.
The GAO offered a qualified verdict on the industry response: “As AI capabilities evolve, AI providers will continue to refine their safeguards.” It added: “However, in its current state, this technology can potentially be dangerous and its misuse represents a growing risk to national and economic security.”
The report also acknowledged the counterweight: “The potential benefits of this technology are immense, but they can also benefit malicious actors.” For Hassan, that trade-off requires congressional scrutiny now, not another cycle of voluntary refinement. January’s hearings will test whether the industry’s claims of model improvement and stronger safeguards match the speed and reach documented by watchdogs.
via fastcompany.com/rss (Source)
More from Priya Raman
Show full bio
Senior reporter covering technology and product news in corporate strategy and chief-executive leadership at CEO Frontier.
450 articles
Adjacent sheets
- GAO Test: AI Chatbots Planned Real Attacks on US Power Grid
- Four AI chatbots produced bombing and power-grid attack plans
- House hearing finds no cyberattack link to recent U.S. outages
- Trump can fight state data center bans with existing federal law
- FAA Ground Stop Triggered by L3Harris Router Flood, Not Cyberattack