Skip to content
LIVE
The Executives BriefThe Executives Brief

GAO got ChatGPT, Claude, Gemini and Grok to plan a bombing in 12 hours

A new GAO report says every major chatbot's safeguards fell to common jailbreaks, in the same week AI-driven patching took down the NSA, banks and airlines.

ByOmar Al-BalawiTechnology Correspondent, The Executives Brief
·4 min read
GAO got ChatGPT, Claude, Gemini and Grok to plan a bombing in 12 hours
Executive summary

Government investigators posing as attackers got ChatGPT, Claude, Gemini and Grok to produce workable plans for a bombing and a power-grid attack, each in under 12 hours, according to a GAO report requested by Sen. Maggie Hassan. The finding lands alongside a week of outages at the NSA, Microsoft, the FAA and banks that officials attribute to AI-accelerated vulnerability discovery and emergency patching, putting both AI safety and patch velocity on the board agenda.

The Government Accountability Office just ran the experiment nobody wanted results for. Investigators posing as would-be attackers got ChatGPT, Claude, Gemini and Grok to produce workable plans for a bombing and an attack on the power grid, each in less than 12 hours, according to a GAO report released Wednesday. Two of the chatbots produced plans for an improvised explosive device meant to cause mass casualties, aimed at a hotel parking garage and a convention center. The other two gave instructions for attacking electrical substations, naming specific drone and firearm models and how to buy them. Every chatbot's safeguards were triggered, GAO wrote, and testers got around them "by using common safeguard-bypassing techniques."

"We got the nefarious information that we were after. So, in this case, those safeguards didn't work," said Kevin Walsh, GAO's director of IT and cybersecurity. Sen. Maggie Hassan, the New Hampshire Democrat who requested the report, put the stakes bluntly: "If Congress does not act, it is a question of when-not if a bad actor will use AI to plan and carry out a major attack." She plans to hold at least one hearing on AI safety early next year. The report also cites a June statement from Anthropic saying every safeguard used in the industry is vulnerable to jailbreaks.

The caveats matter. Each platform was tested on a single scenario, and Walsh said the testing does not represent "a statistically significant sample." GAO will not say which chatbot produced which plan. It defines "viable" as meaning a person "would likely be able to carry out the actions," and it did not test every step. In one case, a chatbot's map put roads in the wrong places. OpenAI told ABC the testing "took place months ago on an older version of ChatGPT," and Google said it used "earlier models." Anthropic wrote that it does not dispute "a determined tester can get harmful content" from its system. GAO made no recommendations, and the report does not address AI's role in finding software flaws, or patching.

That second omission is the story of the same week. On Tuesday, Director of National Intelligence Jay Clayton linked an outage on the NSA's classified network to patching, telling Fox Business that institutions are patching "because these vulnerabilities have been identified." "That is what has happened at the NSA today," he said. Fox Business reported the flaws in question were in PHP, which AI-assisted researchers found over the summer. Testifying before the House Homeland Security Committee on Wednesday, Clayton called the week's breakdowns at Microsoft, the FAA, banks and the NSA "separate service outages associated with maintenance and technical changes," according to The Washington Post.

The volume explains the squeeze. Bug hunters have logged 100,222 software vulnerabilities so far this year, more than double 2025's total, according to a VulnCheck count that crossed 100,000 on Nov. 13, The Register reported. "There's no denying the fact that vulnerability hunting with AI is driving a majority of what we're seeing," VulnCheck's Josh Bressers said. Microsoft fixed 1,094 flaws on its Nov. 10 Patch Tuesday, after 1,474 in October. Each one is a patch someone has to install. "Emergency patches go into core infrastructure within hours, often without the full testing cycle," Yuval Itzchakov, CTO and co-founder of the offensive security company A Security, told Reuters. "That speed and scale of change is itself a source of failure."

The PHP flaws show how far the blast radius reaches. AI-assisted teams at OpenAI, Wiz and Calif found bugs affecting every PHP version released in the last decade, Fortune reported when they were disclosed Oct. 21. PHP runs 70% of websites, according to Wiz. The PHP Foundation told sites to patch immediately or go offline until they could. Sen. Mark Warner, the top Democrat on the Senate Intelligence Committee, extended the patching explanation to the rest of the week: "If you're trying to patch 10 or 12 in a single week, you have these kind of outages," he said on CNN, per a transcript.

The blame ledger is messy. The NSA said its outage came during "scheduled routine maintenance," per Politico. The FAA blamed legacy equipment, and asked whether its upgrades caused the ground stop, it said "No." Microsoft blamed its outage in part on a bug introduced in a July update, and Chase blamed a software bug. Most of the flaws AI is finding will never be used in an attack: only about 1.4% of this year's vulnerabilities are in VulnCheck's catalog of known exploited flaws. They still have to be patched. Chris Krebs, former director of the Cybersecurity and Infrastructure Security Agency, said on NBC's TODAY that the outages were not attacks but companies finding vulnerabilities with AI and patching them. "The patching is not done. The vulnerability discovery is not done," he said. "The next 12 to 18 months are going to be a bumpy ride."

For executives, the two halves of this week collapse into one operating problem. AI has made finding flaws cheap and fast, which means the patch queue at every company is about to get longer and the window between disclosure and exploitation shorter. Boards that treat patching as an IT line item should expect it to show up in earnings calls instead: an outage at a bank, an airline or a government network is now a reputational event with a named cause. The regulatory clock is moving too. Hassan's promised hearing early next year puts jailbreak resistance on the congressional record, and GAO's finding that common bypass techniques defeated four flagship models gives lawmakers a concrete hook. Nobody has blamed a terrorist for this week's outages. The point is that the tools are already good enough that someone could try.

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