Government Watchdog Finds AI Models Generated Specific Terror Attack Plans

Federal investigators successfully bypassed safety protocols on four prominent artificial intelligence platforms, obtaining detailed, feasible plans for potential terrorist attacks in under 12 hours, according to a report released Wednesday by the Government Accountability Office (GAO).
The GAO conducted undercover testing in the spring of 2026, targeting AI assistants from Anthropic, OpenAI, Google, and xAI. Investigators reported that each of the four platforms—Claude, ChatGPT, Gemini, and Grok—generated information that could assist in carrying out attacks against crowded public venues or the nation’s electrical grid. The generated responses reportedly included specific details on potential targets, locations, timing, and methodologies for the attacks.

According to the report, two of the platforms provided plans involving the use of improvised explosive devices intended to cause mass casualties, while the remaining two outlined strategies to damage electrical infrastructure for the purpose of inflicting economic harm. While each platform initially triggered safeguards designed to block or remove harmful content, investigators stated they were eventually able to circumvent those protections.
To evaluate the credibility of the information, GAO investigators performed real-world checks and consulted with experts. The study concluded that the AI-generated plans were sufficiently feasible to aid someone actively planning an attack. The report, which was requested by Sen. Maggie Hassan, D-N.H., serves as an assessment of national and economic security risks associated with the technology.

“The potential benefits of this technology are immense, but they can also benefit malicious actors,” the GAO warned in its findings.
The companies involved responded by emphasizing their ongoing efforts to improve safety. Anthropic stated that it takes the findings seriously and acknowledged that determined users can sometimes bypass safeguards; however, the company questioned aspects of the GAO’s methodology and noted it could not independently verify the results. Google pointed out that the investigators tested earlier versions of its models, adding that it continues to refine its technology. OpenAI stated that it welcomed the opportunity to respond and acknowledged that there is “work that remains” to be done. xAI stated that it “continues to work diligently” to improve safety with every model release.

Senator Hassan addressed the implications of the findings on Wednesday, stating, “The AI public safety threat is no longer theoretical. If Congress does not act, it is a question of when – not if – a bad actor will use AI to plan and carry out a major attack.”