AI Platforms Can Generate Feasible Attack Plans, GAO Report Reveals
An investigation into four prominent AI platforms has revealed their capability to assist users in planning and executing feasible attacks on national security and the U.S. economy. The U.S. Government Accountability Office (GAO) reported that prompts directed at Claude, Grok, ChatGPT, and Gemini resulted in the generation of detailed attack plans in less than 12 hours. These plans included specific targets, locations, timing, attack methods, and strategies to evade detection.
The GAO's findings indicate that while generative AI has been recognized for its efficiency in various tasks, it also poses significant risks when utilized by everyday users. The report states that these advancements enable malicious actors to conduct harmful actions more swiftly and effectively, providing less-skilled individuals with detailed expertise in sensitive areas.
Each of the AI platforms was able to produce attack plans in under 12 hours, although the time required could vary based on the topic and the user's experience. The findings were presented to the Congressional Joint Economic Committee (JEC) in October before being publicly released.
In response to the report, representatives from the AI platforms acknowledged the findings and expressed their commitment to enhancing the safety of their technologies. A spokesperson for OpenAI, the creator of ChatGPT, noted that the tests were conducted using an older model and emphasized the considerable safety improvements made since then. They highlighted the rigorous internal testing and external expert reviews conducted to identify and mitigate misuse.
Verification of AI-Generated Attack Plans
The GAO researchers independently verified that the information provided by the AI platforms was sufficient for a user to plan and execute an attack. In one instance, they consulted a bomb squad commander who confirmed that plans for creating an improvised explosive device would indeed result in an explosion. The researchers also visited suggested locations for acquiring materials and confirmed their availability, noting that with minor adjustments, a malicious actor could inflict significant harm.
Intelligence Reviewed
In another example, an AI platform provided guidance on purchasing materials, cleaning evidence, and even establishing a shell company to avoid detection. While the platforms did implement safety measures, such as denying certain requests and suspending user accounts, the GAO found that these guardrails could be bypassed.
Senator Margaret Hassan, who requested the report, indicated that the Senate Homeland Security and Governmental Affairs Committee would address these AI-related threats in its upcoming meetings. She warned that if Congress does not take action, it is only a matter of time before a malicious actor uses AI to orchestrate a major attack.




