Four of the world's most widely used artificial intelligence platforms provided enough information to draw up attack plans considered viable in under 12 hours each, despite triggering safety mechanisms to block the requests. That is the conclusion of an audit by the Government Accountability Office (GAO), the U.S. Congress's watchdog, made public on November 18, 2026.
The tests were conducted in the U.S. spring of 2026 with ChatGPT, Claude, Gemini and Grok. Investigators posed as ordinary users and tried to obtain information for two types of attack: one involving improvised explosives and another targeting electrical infrastructure.
Two platforms provided plans involving explosives placed in vehicles, with hotels and convention centers as targets. The other two produced information for attacks on electrical substations, one using drones and the other firearms.
The results included specific targets, locations and execution methods. In both explosives scenarios, the systems also indicated exact dates and times. In the case of the power grid, one platform suggested a time of year, while the other did not specify the timing of the attack.
The GAO identified the systems only as A, B, C and D, without revealing which company corresponded to each result.
All platforms initially refused the requests, but investigators managed to bypass the protections using known techniques. During or after the tests, there was also content removal and account suspension.
To verify the results, the GAO conducted research, in-person observations and consultations with experts. The audit considered the plans broadly viable, although it did not carry out attacks or validate every step.
The agency noted that the information used was not necessarily new. The difference was in the chatbots' ability to assemble, in a few hours, guidance that could require weeks or months of conventional research.
The report GAO-27-108814, completed on October 16, was released with sensitive information omitted. Representatives of the four companies generally said they would continue improving their platforms' safety mechanisms.



