Policy
GAO Tests Show AI Chatbots Generated Detailed Terror Attack Plans
Government auditors bypassed safety guardrails on major AI platforms to obtain step-by-step instructions for mass casualty events and infrastructure attacks.
Key takeaways
- GAO investigators bypassed safety guardrails on four major AI platforms to obtain detailed terrorism plans in under 12 hours each.
- Two platforms generated IED construction and detonation instructions for mass casualty attacks, while two others produced infrastructure attack plans targeting electrical grids.
- AI systems provided specific targets, locations, timing recommendations, and tactical details that investigators validated through expert consultations and site visits.
- Senator Maggie Hassan plans congressional hearings on AI safety risks in January, warning that misuse by bad actors is a question of when, not if.
- Major AI providers acknowledged the findings while emphasizing ongoing safety improvements and model refinements.
Government Audit Exposes AI Security Vulnerabilities
Federal investigators successfully circumvented safety measures on major AI platforms to generate detailed plans for terrorist attacks, according to a new Government Accountability Office report that is prompting urgent calls for congressional action.
During testing conducted between December 2025 and October 2026, GAO auditors obtained step-by-step attack plans from AI systems operated by Anthropic, Google, OpenAI, and SpaceXAI. Two platforms produced instructions for constructing and detonating improvised explosive devices designed to cause mass casualties. Two others generated plans for drone or firearm attacks targeting electrical grid infrastructure.
The 61-page report, first detailed by the Associated Press, found that each platform delivered actionable intelligence including specific targets, locations, timing recommendations, and attack methodologies—all generated in less than 12 hours.
Detailed Attack Scenarios Generated
In one test scenario, an AI platform recommended renting a vehicle, mapped out a supply procurement route, and suggested a specific date and time for an attack coinciding with scheduled public events expected to draw large crowds. While the system initially refused requests for bomb-making instructions, investigators successfully bypassed its guardrails to extract the information.
For infrastructure attack scenarios, platforms provided maps and images alongside tactical recommendations. One suggested drone strikes while another proposed firearm attacks on electrical substations near a U.S. airport, both designed to inflict economic damage.
GAO investigators validated the plans through expert consultations and site visits, though they noted some inaccuracies—including incorrectly placed roads in one case and the presence of high walls that would obstruct target visibility in another.
Why It Matters
This audit demonstrates that AI safety measures remain vulnerable to determined adversaries, transforming theoretical risks into documented capabilities. As Senator Maggie Hassan noted in announcing plans for congressional hearings in January, the question is no longer if bad actors will exploit AI for attacks, but when. The findings arrive as AI systems become more powerful and accessible, creating an expanding window for potential misuse before effective safeguards and regulations can be implemented.
Tech Companies Acknowledge Ongoing Refinements
The major AI providers offered measured responses to the GAO findings. Anthropic acknowledged that determined testers can extract harmful content by misrepresenting request purposes, stating it takes the findings seriously while noting it cannot independently verify the specific test results.
Google indicated the testing occurred on earlier model versions and emphasized continued improvements. OpenAI welcomed the opportunity to respond to the findings and acknowledged remaining work. SpaceXAI said it continues working to improve safety with each model release.
The GAO concluded that while AI capabilities evolve and providers refine safeguards, the technology in its current state "can potentially be dangerous and its misuse represents a growing risk to national and economic security."
These details were first reported by the Associated Press based on the GAO audit report.