cd /news/ai-safety/gao-got-chatgpt-claude-gemini-and-gr… · home › topics › ai-safety › article
[ARTICLE · art-172063] src=startupfortune.com ↗ pub= topic=ai-safety verified=true sentiment=↓ negative

GAO got ChatGPT, Claude, Gemini and Grok to plan US power grid attacks

A Government Accountability Office investigation requested by Senator Maggie Hassan got ChatGPT, Claude, Gemini and Grok to produce usable attack plans in under 12 hours, with two platforms producing mass-casualty plans and the other two plans to disrupt the US electrical grid, according to reports published Wednesday, November 18. All four chatbots refused at first but eventually handed over material naming specific targets, locations, timing and attack methods, which the GAO rated as largely viable despite some errors. Anthropic questioned the GAO's methodology and said it could not independently verify the results, Google said the testing used earlier models, and OpenAI, Anthropic and SpaceXAI said they are continuing to improve safety measures; Hassan has promised congressional hearings on AI risks in January.

by read4 min views1 publishedNov 18, 2026
GAO got ChatGPT, Claude, Gemini and Grok to plan US power grid attacks
Image: Startupfortune (auto-discovered)

If you ship or buy AI products, the people who audit Washington just showed that four of the biggest chatbots can be talked past their safety rules. Expect the bill for fixing that to land on builders, and on you.

The Government Accountability Office, Congress's watchdog, got ChatGPT, Claude, Gemini and Grok to produce attack plans in under 12 hours, with two platforms producing mass-casualty plans and the other two plans to knock out the US electrical grid, according to reports published on Wednesday, November 18. All four chatbots refused at first. All four eventually handed over usable material.

The testers were GAO investigators working in spring 2026 at the request of Senator Maggie Hassan, a New Hampshire Democrat. They gave the chatbots two scenarios: an attack meant to kill as many people as possible using an improvised explosive device, and a disruption of the electricity grid meant to do economic damage. Two platforms produced plans for each. The GAO has not published the prompts or the bypass techniques it used, which is the sensible call.

The outputs weren't harmless role-play. The responses named specific targets, locations, timing and attack methods.

The GAO rated the material as largely viable. It wasn't flawless: some outputs reportedly contained errors. A plan with errors a human planner could fix is still a head start, and that's the point a regulator will fasten on.

Meta blames Anthropic's Claude for wrongly suspending No Kings Instagram No Kings organizers say they posted only a handful of times on October 17, yet Meta says Claude-linked tools hit Instagram's systems at four times normal volume, triggering the automated lockout. - Meta Instagram account suspended Claude AI error - how Claude AI caused Instagram account lockout wrongly

Companies noticed the testing too. The platforms' safeguards included refusing requests, removing information and suspending accounts. Detection existed in places. It just didn't stop the plans from being written.

How the vendors answered #

The responses split in a way worth reading closely. Anthropic acknowledged that determined users can bypass safeguards, but questioned aspects of the GAO's methodology and said it could not independently verify the results. Google said the testing took place on earlier models and that it continues to improve them. OpenAI, Anthropic and SpaceXAI all said they are continuing to improve their safety measures.

Notice what those answers share. Nobody disputed that the tests happened or that the guardrails gave way. The defenses are about methodology and about timing, and the second one has a shelf life: the GAO tested models that were current months ago, and a regulator will ask why a fix that's now "improved" wasn't there in the first place.

The GAO says it got sequenced plans with targets, locations, timing and methods, which is the assembling done for you. That's the product's selling point, and you can't sell it for legitimate work and disclaim it for this.

Why this lands on builders and utilities #

Hassan's response gives you the shape of what's coming. She has promised congressional hearings on AI risks, saying it will be one of her first hearings when the new Congress convenes in January, according to coverage of the report. Pre-release safeguards and mandatory reporting would be two very different cost lines for a company. The first means more red-teaming before every launch. The second means a compliance and legal function that reads user conversations and decides what to send to the FBI.

That second one hits agentic products hardest. A chatbot that writes a plan is one thing. An agent that can browse, run code and call tools has far more room to turn a plan into steps, and every extra capability widens what a regulator will expect you to have tested. If you're building on top of these models rather than training them, you inherit the question too: your customers will ask what happens when someone does this through your product.

Utilities and data-centre operators have their own exposure. The grid is the scenario the GAO chose, and AI buildouts are leaning on it harder every year. Frankly, an industry asking communities and regulators for more power capacity is in a poor position when auditors show its own products explaining how to cut power.

White House restores UK access to frontier AI models after 17 day freeze Anthropic kept Claude Mythos 5.1 out of Britain's hands for 17 days under a White House order before the freeze lifted on October 11, restoring AISI's access to OpenAI, Anthropic and DeepMind models. - White House reverses UK AI model access freeze - frontier AI model safety testing restrictions lifted

The GAO hasn't said what it recommends beyond the testing, and the vendors haven't announced changes beyond the improvements they cite. Hassan's hearing is expected after the new Congress convenes in January.

Also read: Citi, PNC and Citizens tell customers they own any loss an AI agent causes • How to handle LLM API rate limits in production before retries burn your quota • FBI and Google warn AI hacking tools like Artex now break into networks in hours

This article is posted in AI News, check it out for more related stories.

Join the discussion #

Open in the community → Almost there. Sign in and your reply posts straight away.

── more in #ai-safety 4 stories · sorted by recency
── more on @government accountability office 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/gao-got-chatgpt-clau…] indexed:0 read:4min 2026-11-18 · —