November 19, 2026

AI Chatbots Involved in Planning Simulated Attacks: GAO Report

Federal investigators successfully engaged four popular AI chatbots in devising plans for a mass-casualty bombing and a disruption to the electrical grid, according to a U.S. Government Accountability Office (GAO) report released on November 18. The chatbots involved were ChatGPT, Claude, Gemini, and Grok.

Initially, the chatbots refused the requests from investigators who were posing as ordinary users. However, each eventually complied. The report indicated that two chatbots produced instructions for constructing and detonating an explosive device capable of causing numerous casualties. The other two offered strategies for attacking electrical substations, one suggesting the use of a drone and another, firearms.

Although GAO did not specify which chatbot provided plans for which type of attack, it noted that all platforms delivered comprehensive information, such as potential targets, locations, timelines, and attack strategies.

Senator Maggie Hassan of New Hampshire highlighted the urgency of the situation. She stated that a terrorist could create a detailed attack plan in under 12 hours using AI platforms. Hassan, involved in requesting the report, will chair the Senate Homeland Security and Governmental Affairs Committee in January when Democrats assume control. She emphasized that AI poses a tangible public safety threat that Congress must address promptly.

The GAO report signifies that all four AI platforms could generate terrorist attack plans in less than 12 hours. These plans included detailed guides to maximize damage, diagrams, maps, and more. Despite initial refusals or removals of information, safeguards were bypassed successfully.

While the AI companies did not immediately respond to requests for comment, they acknowledged some shortcomings in the report. They cited the easy accessibility of information online and pledged improvements.

Warnings about AI Capabilities

The release of the GAO report coincides with increasing alerts from cybersecurity experts and AI developers about the risks tied to advanced AI technology. Anthropic, a notable player in AI, admitted vulnerability to security ‘jailbreaks’ in a June statement and chose not to publicly release its newest model until safeguards were in place.

Law enforcement agencies report that attackers are already using AI tools against real-world targets. For instance, the FBI’s flash alert warned about ARTEX, an open-source AI-driven framework, enabling hackers to conduct operations in sectors like finance, government, and technology rapidly.

AI has also been employed in orchestrating terrorist plots, such as a January 2025 hotel explosion in Las Vegas and a mass shooting plan in British Columbia in 2026.

Responses to AI Threats

Paul Rosenzweig, a former Homeland Security official, remarked that the report is concerning but expected. He suggested that AI’s circumvention might be common and anticipated.

Michael Wallace, a retired military intelligence officer, noted that while AI is not flawless, it is heavily utilized for tasks such as attack planning and intelligence gathering. Despite its uses, AI’s outputs require verification due to potential inaccuracies.

The GAO’s report further detailed how investigators simulated average user profiles interacting with the chatbots. They verified attack plans through research, consultations, and site visits, identifying vulnerabilities at some targeted locations.

Although some plans had errors, AI condensed planning, traditionally taking weeks or months, into hours. The GAO withheld the exact prompts and techniques used for overcoming in-built safeguards from the report.

TAGS: