GAO test finds chatbot safeguards failed to block attack planning
Investigators with the Government Accountability Office used Claude, Grok, ChatGPT and Gemini to produce information that could assist malicious actors in planning and executing attacks within 12 hours. The congressional watchdog said well-known techniques bypassed the platforms’ safeguards, yielding feasible plans involving homemade explosives and attacks on the U.S. electrical grid.
The tests covered explosive attacks and electrical infrastructure sabotage. Outputs included target selection, equipment acquisition and ways to evade detection. Investigators checked the information through research, surveillance and consultation with a bomb squad commander. Each platform was tested once, with scenarios randomly assigned. The GAO withheld model-specific results and operational details to prevent misuse.
The companies told the GAO they continually test their systems and had strengthened safeguards since the experiments. OpenAI said the testing involved an older ChatGPT version and pointed to subsequent safety improvements. Some platforms suspended investigators’ accounts after the study, but the GAO did not disclose whether testing prompted any law enforcement response.
Sen. Maggie Hassan, who requested the study, plans hearings and legislation to strengthen AI safety requirements. She warned that competitive pressure to release models quickly makes enforceable safeguards necessary. The GAO concluded that misuse of the technology represents a growing national risk.