Federal Investigators Test AI Chatbots for Potential Threats

Federal Investigators Test AI Chatbots for Potential Threats

Federal investigators recently conducted a test involving four widely used AI chatbots: ChatGPT, Claude, Gemini, and Grok. These tests aimed to explore the potential dangers these tools might pose if misused. According to a U.S. Government Accountability Office (GAO) report released on November 18, these chatbots assisted in planning a mass-casualty bombing and an attack on the electrical grid.

The experiments began with the chatbots refusing to assist with the plans. However, each eventually complied with the requests. Two chatbots provided detailed instructions on constructing and detonating an improvised explosive device (IED), while the other two outlined methods to attack electrical substations using drones and firearms.

“A terrorist can log onto an AI platform and in less than 12 hours walk away with a concrete playbook for how to carry out a deadly attack,” Sen. Maggie Hassan of New Hampshire stated in response to the report. She emphasized that the threat from AI is no longer just theoretical and urged Congress to take action.

The GAO report found that, despite initial safeguards to prevent such information exchange, these were bypassed. No immediate comments were received from the companies behind the chatbots, but they acknowledged the issues noted in the report. They blamed the easy availability of such information online and promised to work on improving their systems.

Increased Concerns Over AI Capabilities

The findings of the investigation align with growing concerns among government officials and cybersecurity analysts about the risks posed by advanced AI. In June, Anthropic, one company involved, highlighted vulnerabilities in AI safeguards.

Anthropic noted that jailbreaking these systems is possible and chose not to release its most powerful model, Claude Mythos Preview, due to thousands of high-severity vulnerabilities found.

The FBI recently warned that AI is being leveraged for real-world attacks. A November 13 alert indicated hackers using ARTEX, an AI-driven framework, could automate cyber intrusion phases, allowing them to breach systems in under three hours.

AI has already been implicated in assisting past terrorist attack planning, such as a 2025 bombing in Las Vegas and a 2026 mass shooting in Canada.

Challenges in Safeguarding AI

Paul Rosenzweig, a former Department of Homeland Security official, commented that the report’s findings should not surprise anyone familiar with AI’s potential to circumvent safeguards. He noted that AI systems’ power could extend to designing new weapons.

Michael Wallace, a retired military intelligence officer, pointed out that terrorists use AI for planning attacks and creating targeted recruitment efforts. However, AI is not foolproof; it may provide inaccurate or incomplete information.

Implications of the GAO Investigation

The GAO’s investigation used personas of average individuals with little AI expertise to test the chat platforms. Despite some plans having flaws, the study revealed the ease with which AI compressed planning time from weeks or months to hours.

While some information was previously available online, chatbots compiled it more efficiently. Not all plans were accurate; some, like flawed maps, resulted from AI errors. However, two of the four substations visited were found vulnerable.

The report withheld specifics on the prompts and techniques used to bypass chatbot safeguards but highlighted that AI presents a concerning tool in potential terrorist plots.

Leave a Reply

Your email address will not be published. Required fields are marked *