US GAO covert investigation: 4 major AI services including ChatGPT all generated information usable for attack planning in under 12 hours
ChatGPTら主要AIサービス4種、12時間未満で攻撃計画の情報を生成―米GAOの覆面調査 | 財経新聞
The US Government Accountability Office (GAO) published a covert investigation report on November 18; investigators bypassed the safety protections of four AI services—ChatGPT, Claude, Gemini, and Grok—and all four provided, in under 12 hours, information usable for planning a major attack within the United States. In tests in spring 2026, two of the services generated information related to a bombing attack in a major US city, and the other two generated attack information targeting the US power grid; after review, the GAO considered the information largely feasible but containing geographic detail errors, and no real attack was carried out during the tests. The GAO acknowledged that most of the basic information was already public, but that AI could organize it into a plan within hours, potentially accelerating investigations that would otherwise take weeks to months. Both OpenAI and Anthropic disputed the investigation methodology; the latter said the number of tests was limited and reflected its system as of May; the GAO emphasized that this was a limited assessment at a specific point in time. The investigation was conducted at the request of Senator Maggie Hassan, who is considering legislation requiring companies to conduct stricter safety testing before releasing new models.
In the report, the GAO responded to the methodological objections of OpenAI and Anthropic and explained the assessment's limitations, so that readers can calibrate the scope of applicability of the conclusions accordingly.