The Government Accountability Office (GAO), the investigative arm of the US Congress, released a report on November 18 at the request of Sen. Maggie Hassan of New Hampshire. Investigators posing as would-be attackers tested ChatGPT, Claude, Gemini and Grok on one scenario each. According to Fortune, every chatbot's safeguards triggered, but testers got around them with common techniques. GAO says each chatbot then produced material it considers viable for violent attacks on public sites and critical infrastructure, although some details were wrong.
GAO notes that the sample was not statistically significant, and it made no recommendations. OpenAI says the testing used an older ChatGPT version, and Google says it involved earlier models. Anthropic says it does not dispute that a determined tester can obtain harmful content. For companies that deploy these chatbots, the findings show vendor safeguards can fail. Hassan plans a hearing on AI safety early next year.
