GAO testers got ChatGPT, Claude, Gemini and Grok to draft viable attack plans in under 12 hours each — safeguards fired, then were bypassed
The US Government Accountability Office publicly released a report on November 18 finding that investigators posing as ordinary users got each of the four most-used chatbots — ChatGPT, Claude, Gemini and Grok — to produce a viable attack plan in under 12 hours. Two platforms produced plans for a mass-casualty explosive device and the other two produced plans for attacking electrical substations; GAO does not say which chatbot produced which plan.
Every platform's safeguards fired first — refusals, stripped content and account action — but investigators got past them using what GAO calls common safeguard-bypassing techniques. GAO defines “viable” as a plan a person would likely be able to carry out, while noting the information was largely not novel and some details were wrong, including misplaced roads on one map.
The tests carry stated limits: each platform faced a single scenario in spring 2026 testing, on older model versions — Claude 4, Grok 4, the GPT-5 series and Gemini 3 Flash — which GAO itself calls not a statistically significant sample. OpenAI and Google said the tests used outdated models, Anthropic said it did not dispute that a determined tester could elicit harmful content, and Sen. Maggie Hassan, who requested the report, says AI safety will be an early priority and hearing topic when the new Congress convenes in January.