In tests conducted in spring 2026, the U.S. Government Accountability Office (GAO) found that four AI platforms each produced information it judged could help with attack planning in less than 12 hours. The platforms were ChatGPT, Claude, Gemini and Grok. GAO’s report is dated October 16; MeriTalk reported that it became public on November 18.
What the spring 2026 tests found
GAO selected the four platforms using usage-survey and U.S. web-traffic data from April 2026. Investigators assigned two platforms at random to a scenario involving an improvised explosive device and two to a scenario involving an attack on the electrical grid. The report does not identify which platform received which scenario.
All four platforms initially triggered safeguards, including refusals. Investigators later bypassed safeguards on each one and obtained information GAO judged feasible to use in attack planning. Each platform produced that information in under 12 hours; GAO noted that the time could vary by topic and user experience. Investigators checked selected information through research, surveillance, site visits and expert interviews. They did not build an improvised explosive device or carry out an attack. GAO’s report describes the tests and their findings.
What GAO meant by “viable”
GAO used “viable” to describe information with which a person would likely be able to carry out a plan. That is a finding about the information’s feasibility, not a prediction that an attack would succeed. GAO said success would also depend on outside factors such as traffic, weather and law enforcement.
A feasibility finding from a limited test is not a safety ranking of the platforms.
Why the findings do not rank the chatbots
Each platform faced just one randomly assigned scenario, and GAO said its sample was not statistically significant. The agency said the results do not support comparisons among the platforms. Because the scenario assignments are anonymized, the report also cannot show which named chatbot produced information for either scenario.
The findings describe systems as tested in spring 2026. GAO characterized the results as a point-in-time assessment, rather than a measure of later versions.
The reported policy response
After the report’s release, Sen. Margaret Wood Hassan called for AI safety legislation and said one of the next Congress’s first hearings would examine risks to homeland security and critical infrastructure, MeriTalk reported. GAO issued no recommendations in the report.