US watchdog obtains attack-planning information from four AI chatbots
A congressional watchdog says researchers bypassed safeguards on four popular AI platforms to obtain information for attack plans. The small test did not measure how often such attempts would succeed.

Review information is loading.
The US Government Accountability Office (GAO) says its investigators obtained information for attack plans from ChatGPT, Claude, Gemini and Grok after finding ways around their safety safeguards.
The congressional watchdog tested the four platforms in May at the request of Democratic senator Maggie Hassan. Two were randomly assigned a scenario aimed at causing mass casualties; the other two were given a scenario involving economic damage. Each provided information for a plan in less than 12 hours, the GAO said.
The investigators’ initial requests did not produce harmful information, and all four platforms refused some requests during the tests. Researchers subsequently got around the safeguards. The report does not disclose the methods used or identify which platform produced each result.
The GAO assessed the information through research, site visits and discussions with experts, and found it was largely useful for planning the scenarios. It also found errors: one proposed date did not coincide with an event, while high walls obstructed views at some locations. Investigators did not build an explosive device or safely verify every step.
The test was small and does not establish how often similar requests would succeed. Nor was it designed to measure whether chatbots made the information quicker or easier to obtain than other sources. Investigators began with free accounts and also used paid subscriptions for Claude, Gemini and Grok.
According to the GAO’s summary of company responses, OpenAI said it was not always clear which parts of the plans came from ChatGPT and which came from researchers or other sources. The watchdog said its researchers used their experience and consulted outside sources. Anthropic pointed to the small sample and lack of a control group, and said it had added safeguards since the test.
Hassan said she wants rules requiring safety measures before new models are released, reporting of illegal activity to law enforcement and regular updates to safeguards. She plans to hold at least one Senate hearing on AI safety in early 2027.
Sign in to give this article a thumbs up or down.
From the discussion
Private test discussion. Comments are readers’ views and are not yet automatically fact-checked. Editing is available for 60 seconds after posting.
Sign in with a confirmed reader account and choose a username to read comments and participate.
Sorting applies to top-level comments; replies remain oldest first. New comments and likes can change the order. Refresh for the current ranking.
Loading comments…