US Watchdog Exposes Security Vulnerabilities in Leading AI Chatbot Platforms
A US GAO report reveals that leading AI chatbots by OpenAI, Anthropic, Google, and xAI can be manipulated to provide actionable data for planning attacks within hours.

US Government Accountability Office (GAO) investigators conducted a covert experiment testing leading artificial intelligence platforms, revealing that major chatbots can eventually be manipulated into providing actionable information for planning mass casualty attacks and critical infrastructure sabotage. The findings were detailed in a report published by the congressional watchdog.
Covert Experiment and Platform Testing
The investigation evaluated four prominent AI platforms: OpenAI's ChatGPT, Anthropic's Claude, Google's Gemini, and xAI's Grok. Two systems were tested against mass casualty terrorist attack scenarios in major US cities, while two others faced scenarios targeting the US electrical grid. Initially, all four platforms activated safety protocols and rejected the dangerous prompts. However, as investigators continued the conversational dialogue, every system ultimately provided data that could assist in executing attacks.
According to the GAO, fewer than 12 hours of interaction were required in each test to obtain information facilitating a practical operational plan. The report deliberately withheld which specific chatbot was paired with each scenario to prevent the exploitation of its findings. The investigation was initiated at the request of Democratic Senator Maggie Hassan of New Hampshire, who warned that the security threat posed by generative AI is no longer theoretical.
Operational Details and Field Verification
In the mass casualty scenarios, the AI systems suggested targets, optimal timing, and methodologies, alongside required preparations. One model proposed targeting a hotel parking garage and generated an image of the location, while another recommended attacking a convention center accompanied by a detailed preparation schedule. For the electrical grid scenarios, chatbots suggested drone and firearm attacks, providing specific facility details, maps, and visual illustrations.
Investigators verified the information by conducting field visits to the mentioned sites and consulting subject matter experts, including an explosive ordnance disposal commander and an electrical infrastructure specialist. While the experts confirmed the intelligence was sufficiently practical to aid malicious actors, they also identified inaccuracies, such as proposed event dates that did not align with convention schedules and mispositioned roadway maps.
Speed and Industry Response
The central concern raised by investigators was not the novelty of the information—much of which remains accessible online—but the unprecedented speed with which artificial intelligence aggregated data, correlated disparate details, and structured actionable plans within hours instead of months. The GAO acknowledged that the study did not constitute a broad statistical sample, and noted that investigators applied their own professional expertise alongside the AI tools.
Representing the developers, Anthropic stated that the evaluation tested versions available in spring 2026 and that defensive measures have since been upgraded. Google emphasized its ongoing efforts to enhance Gemini safety filters, while OpenAI highlighted that ChatGPT rejected many hazardous requests during the evaluation and questioned the distinction between system-generated data and external investigator inputs. xAI pledged continued improvements to its safety and oversight mechanisms.





