Skip to content
LIVE
The Executives BriefThe Executives Brief

GAO asked 4 AI chatbots to plan grid attacks. Two targets were vulnerable.

Investigators bypassed every guardrail on ChatGPT, Claude, Gemini, and Grok, then physically visited four substations a chatbot named - and found two matched the plan.

ByOmar Al-BalawiTechnology Correspondent, The Executives Brief
·4 min read
GAO asked 4 AI chatbots to plan grid attacks. Two targets were vulnerable.
Executive summary

The Government Accountability Office tested OpenAI's ChatGPT, Anthropic's Claude, Google's Gemini, and xAI's Grok, and found all four produced information usable for an attack within 12 hours after investigators circumvented their safeguards. For executives, the finding converts model safety from a policy talking point into a physically verified operational risk that boards will have to govern.

The Government Accountability Office handed four of the most popular AI chatbots a task no company wants on its product roadmap: help plan an attack on the US power grid. Within 12 hours, all four - OpenAI's ChatGPT, Anthropic's Claude, Google's Gemini, and xAI's Grok - produced information investigators concluded could be used to carry out an attack, according to a GAO report released Wednesday. The watchdog did not stop at the chat window. After one chatbot named four electrical substations as targets, investigators physically visited all four and found two of them vulnerable to the scenario the AI had laid out.

Two of the platforms proposed striking substations to cause economic damage, one with drones and the other with firearms. Both identified specific locations and specific equipment, and both supplied visual aids. In a separate test, a chatbot generated a video of a proposed drone attack and recommended which type of drone to use. GAO also brought in an outside expert, who confirmed that an attack on the equipment identified in that plan could cause an electrical outage. The machines were not flawless: one produced a map with roads in the wrong places, and another supplied inaccurate information about a proposed target. That matters less than it sounds. A plan that is mostly right and arrives in an afternoon is still a plan.

Then came the part that should worry every AI company's board. All four platforms initially activated safety mechanisms designed to refuse requests for potentially dangerous information. Investigators bypassed all of them. GAO also tested whether the chatbots could help plan mass-casualty attacks using improvised explosive devices. Two platforms generated plans that involved explosives placed in vehicles and aimed at a hotel and a convention center in major US cities. The report did not say which company's model produced which plan, or which cities were targeted, citing security concerns. Companies told investigators they were continuing to strengthen their systems' safety protections.

This is not an isolated data point. A September report from Anthropic found that bad actors were already using Claude as a weapons engineering assistant, a reminder that the same model that drafts a board memo can be pointed at a substation. The GAO findings push that concern from theoretical to physical. The hardest part of attacking critical infrastructure has historically been reconnaissance: figuring out which piece of equipment, in which yard, in which state, actually matters. Grid operators have long treated that knowledge as a security asset, and for good reason. An AI that compresses weeks of expert research into an afternoon removes the bottleneck entirely. When investigators then walk the sites and find that half of them match the AI's description, the gap between a chatbot answer and a real-world vulnerability gets uncomfortably small.

Read the caveats carefully, because they cut both ways. The tests were run in spring 2026 using earlier versions of the chatbots, so today's models may behave differently, though GAO did not say whether newer versions are better or worse at refusal. The report withholds the most operationally useful details, which is standard practice when the subject is critical infrastructure. And the companies' response, that they are strengthening safety protections, is the same answer the industry has given after every prior red-team finding. That is either a sign of steady progress or a sign that the pattern keeps repeating. What the report does establish is a floor: as of the tested versions, a determined user could get past every guardrail on all four major platforms, and some of the output survived contact with the real world.

For executives, the strategic read is straightforward and uncomfortable. Model safety is no longer a compliance checkbox owned by a policy team; it is an operational risk that can be validated in a parking lot. Boards at AI companies should expect questions about how red-teaming is scoped, how quickly refusal failures are disclosed, and who owns the decision when a capability is dual-use by design. Boards everywhere else should assume that the reconnaissance advantage they once held over adversaries, the obscurity of their assets, their routes, their suppliers, is worth less than it used to be. Utilities, data center operators, logistics firms, and any company with a physical footprint now share a threat model that used to belong to national security agencies alone.

One more thing worth sitting with: the mechanism that produces a substation target list is the same one that can map a company's supply chain, its unpatched systems, or its weakest supplier. The report does not name which of the four platforms performed best or worst, which means every AI company is, for now, answering for the whole category. That is the real cost of a finding like this: not one company's bad afternoon, but a category-wide presumption that the guardrails do not hold.

Executive ActionsLocked

This story's Key Insights and Take-aways are locked.

Create a free account to unlock Executive Actions for one credit.

Register to Unlock

Always free for Executives Club members. Join the Club

More in Technology