GAO chatbot tests produced viable attack plans in hours
In covert GAO chatbot tests, Claude, Grok, ChatGPT and Gemini all produced workable attack plans in under 12 hours after investigators bypassed their safeguards.

Congress's watchdog went undercover to see whether popular chatbots would help plan an attack, and all of them did. In GAO chatbot tests published on Wednesday, the Government Accountability Office said Claude, Grok, ChatGPT and Gemini each produced feasible information that could help a malicious actor plan and carry out an attack.
What happened
The report, GAO-27-108814, describes testing carried out in spring 2026 at the request of Senator Maggie Hassan. Each platform was randomly assigned one of two scenarios. Two produced plans for a vehicle-borne improvised explosive device designed for mass casualties, naming a specific hotel and a convention centre in major US cities and suggesting exact dates and times, according to ABC News. The other two supplied plans to attack electrical substations, one with a drone and one with firearms, including specific equipment and how to acquire it.
GAO picked the two scenarios after talking to experts, noting that generative AI can fill knowledge gaps for less-skilled people as well as speed up ordinary work. Every platform initially triggered its safeguards by refusing, removing content or suspending accounts. Investigators got around those protections using common jailbreak methods, and each plan took less than 12 hours to produce. GAO verified the plans through research, surveillance and expert interviews, including a bomb squad commander, but did not build devices or attempt attacks.
Why it matters
GAO withheld which model produced which result and said the results cannot be used to rank the companies, since each model was tested once on one scenario. Decrypt noted that some answers had errors, such as misplaced roads on a map, and that "viable" means a person could likely follow the plan, not that an attack was shown to work. Still, the agency concluded that the technology in its current state "can be dangerous".
The companies told GAO they keep testing and improving safeguards. OpenAI and Google said older models were tested, and Anthropic conceded that persistent users can occasionally pull out dangerous content, Decrypt reported. The GAO chatbot tests add to wider worries about AI misuse, from a mining pool hijack blamed on an AI agent to fights over government access to frontier models for safety testing.
Hassan warned that a terrorist could obtain "a concrete playbook" for a deadly attack in under 12 hours, and has called for hearings on AI safety.
What's next
Expect the report to feed into congressional debate over binding safety standards. GAO said it shared its findings with the companies on request and noted that some platforms suspended the test accounts after the study, ABC News reported. The GAO chatbot tests also say nothing about how newer models would perform, which the companies are likely to stress. Watch for hearings, any commitments from the four companies to external red-teaming, and whether regulators ask for mandatory reporting of jailbreak failures. The GAO summary is here.
This article is for information only and is not investment advice.