A U.S. Government Accountability Office audit found that all four major AI platforms it tested produced information assessed as capable of supporting attack planning. The systems generated details covering specific targets, locations, timing, and attack methods, according to GAO findings summarized in a release from Sen. Maggie Hassan’s office.
The audit ran from December 2025 through October 2026, with hands-on testing conducted in spring 2026. The GAO did not publicly identify which platform produced which outputs, limiting the ability to attribute individual failures from the available record.
How the Tests Were Structured
Investigators tested two attack scenarios, encountered initial refusals, and ultimately bypassed the platforms’ safeguards.
Testers worked through two categories of scenarios: planning a mass-casualty event using an improvised explosive device, and attacking electrical infrastructure in major U.S. cities using drones or firearms. Investigators initially encountered refusals and deleted responses. They ultimately bypassed those safeguards using techniques the GAO declined to detail publicly.
Independent experts reviewed the generated content and assessed it as feasible and capable of causing harm. The GAO verified viability through research, surveillance, and expert interviews.
The Findings in Plain Language
The GAO’s own words describe outputs from all four systems without exception.
“Each of these platforms provided detailed information that could be used to carry out attacks, such as specific targets, locations, timing, and attack methods.” Government Accountability Office.
Sen. Hassan characterized the findings directly: “A terrorist can log onto an AI platform and in less than 12 hours walk away with a concrete playbook for how to carry out a deadly attack.” She has said she plans to pursue legislation addressing AI security risks and to hold hearings focused on national security and critical infrastructure threats.
What the Companies Said
AI developers say their newer models carry stronger protections, but no independent testing has confirmed that those gaps are closed.
The four companies told investigators that the testing involved older versions of their systems and that they have since strengthened their safety protections. That response is a mitigation claim; the available reporting does not identify independent testing that confirms current deployments have closed the gaps the audit identified.
The companies’ response also does not address the core finding. As deployed during the testing window, publicly available systems could be manipulated into producing attack-related guidance, according to the GAO’s findings.
Anthropic’s Separate Disclosure
Beyond the audit, Anthropic documented six company-reported misuse cases involving weapons-related software.
Separately from the GAO investigation, Anthropic published threat-intelligence findings identifying six cases of Claude being misused in connection with software related to conventional weapons, including firearms, missiles, armed drones, bombs, and associated targeting systems. The company traced those actors to China, Russia, and Yemen.
Anthropic’s disclosure is a self-reported finding, not independently adjudicated evidence. It nonetheless adds context to the audit by suggesting the concern extends beyond controlled testing environments.
The Accountability Gap That Remains
One audit is not a safety program, and company assurances are not a substitute for independent verification.
The GAO findings raise questions about whether safety evaluations conducted before a model’s release are sufficient when adversarial prompting can bypass protections after deployment. Recurring, independent testing across current model versions and across paid and free tiers alike would offer more reliable assurance than one-time pre-launch reviews, though the GAO itself stopped short of prescribing that remedy.
Mandatory reporting of credible misuse attempts is a policy gap the audit makes visible. Until companies face clear obligations to disclose known failures, those failures may go unreported. Without independent regulatory authority to verify safety claims, the gap between what developers say their systems can resist and what a determined user can extract will remain largely untested.






























