Brief IA
›

The GAO Obtains Attack Plans for Four Chatbots

Models & LLM·Tom Levy·

The GAO Obtains Attack Plans for Four Chatbots

The GAO Obtains Attack Plans for Four Chatbots
Key Takeaways
1The GAO tested ChatGPT, Claude, Gemini, and Grok in the spring of 2026.
2Investigators deemed two of the four inspected substations vulnerable.
3The tests involved earlier versions of the systems, and some responses were incorrect.
Why it matters : Investigators obtained information deemed actionable despite the initial protections of the four platforms.

The Government Accountability Office (GAO), a U.S. public oversight agency, obtained information from four chatbots in 12 hours that its investigators deemed usable for planning an attack. The tests focused on U.S. infrastructures as well as attacks likely to cause numerous casualties.

Previous Versions and Erroneous Responses

The evaluation campaign dates back to spring 2026 and involved earlier versions of the systems. For security reasons, the GAO does not specify which platform was responsible for each scenario or which cities were considered as targets.

Some responses contained errors: one map incorrectly located roads, while another chatbot provided inaccurate information regarding a potential target. Companies informed investigators that they were continuing their efforts to strengthen protections.

Two Substations Found Vulnerable After Inspection

To assess the credibility of the proposals, investigators combined in-depth research, expert consultations, and inspections of the designated sites. The same chatbot selected four substations: all were visited, and two were deemed vulnerable to the proposed attacks. An expert consulted confirmed that striking the designated equipment in a plan could lead to a power outage.

La lettre du soirGratuit · 19 h

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

The scenarios concerning the grid came from two chatbots and aimed at economic damage. One suggested drones, while the other proposed firearms; both specified sites, equipment to target, and provided illustrations. In another test, one platform even produced a video depicting the proposed attack with a drone, accompanied by a recommendation on the type of device.

Protections Bypassed and Attacks Tested

ChatGPT, developed by OpenAI, Claude by Anthropic, Gemini by Google, and Grok by xAI all initially employed protections designed to deny dangerous requests. However, investigators managed to bypass them. The requests were not limited to the power grid: the GAO also assessed the preparation of attacks using improvised explosive devices.

Two platforms thus proposed scenarios using vehicles loaded with explosives against a hotel and a convention center located in major U.S. cities. In September, Anthropic indicated that malicious actors were using Claude to assist them in weapon engineering.

Brief IA — L'actualité IA en français

L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.