
Washington convinces 4 AI chatbots to plan terrorist attacks as faulty internet patches undermine banks and airlines.
Government investigators posing as potential attackers led ChatGPT, Claude, Gemini and Grok to develop feasible plans for a bombing and an attack on the power grid, each in less than 12 hours, according to a Government Accountability Office report released Wednesday.
The report came at the end of a week in which AI has already been linked to real damage, although no one blamed a terrorist. On Tuesday, Director of National Intelligence Jay Clayton linked an outage on the NSA’s classified network to patching, saying institutions are installing patches “because these vulnerabilities have been identified,” on Fox Business. “This is what happened at the NSA today,” he said. Fox Business reported that the flaws in question involved PHP, which AI-assisted researchers discovered over the summer. Testifying before the House Homeland Security Committee on Wednesday, Clayton characterized the week’s outages at Microsoft, the FAA, banks and the NSA as “separate service outages associated with maintenance and technical changes,” according to the Washington Post.
“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,” said Sen. Maggie Hassan, a New Hampshire Democrat who requested the report, according to Nextgov. She plans to hold at least one hearing on AI safety early next year.
What the GAO found
Each chatbot’s protections were triggered and testers bypassed them “using common protection bypass techniques,” the GAO wrote. Two of the chatbots produced plans for an improvised explosive device intended to cause mass casualties, targeting a hotel parking lot and a convention center. The other two gave instructions for attacking electrical substations, naming specific models of drones and guns and how to purchase them.
“We obtained the harmful information that we were looking for. So in this case, these protective measures did not work,” said Kevin Walsh, GAO’s director of information technology and cybersecurity. The report cites a June statement in which Anthropic said every protection used in the industry was vulnerable to jailbreaks.
Warnings matter. Each platform was tested on a single scenario, and Walsh said the tests do not represent “a statistically significant sample.” The GAO won’t say which chatbot produced which plan. It defines “viable” to mean that a person “would probably be able to carry out the actions” and has not tested every step. The GAO found that some details were wrong: In one case, a chatbot’s map showed roads in the wrong places. OpenAI told ABC that testing “took place months ago on an older version of ChatGPT,” and Google said it used “prior models.” Anthropic wrote that it does not dispute that “a determined tester could obtain harmful content” from its system.
GAO made no recommendations. The report does not address the role of AI in finding software vulnerabilities or applying patches.
The damage that occurred
Bug hunters have recorded 100,222 software vulnerabilities so far this year, more than double the 2025 total, according to a VulnCheck tally that passed 100,000 on Nov. 13, The Register reported. “There’s no denying that AI vulnerability scanning is driving the majority of what we’re seeing,” said VulnCheck’s Josh Bressers. Microsoft fixed 1,094 flaws during its patch Tuesday, November 10, after 1,474 in October.
Each is a patch that someone needs to install. “Emergency patches are introduced to the core infrastructure within hours, often without the full testing cycle,” Yuval Itzchakov, CTO and co-founder of offensive security company A Security, told Reuters. “This speed and scale of change is in itself a source of failure. »
PHP flaws show the pressure. AI-assisted teams from OpenAI, Wiz and Calif discovered bugs affecting every version of PHP released over the past decade, Fortune reported when they disclosed on October 21. PHP powers 70% of websites, according to Wiz. The PHP Foundation has asked sites to update immediately or go offline until they can.
Sen. Mark Warner, the top Democrat on the Senate Intelligence Committee, extended the explanation of the fixes to the rest of the week. “If you try to release 10 or 12 patches in a single week, you run into these kinds of outages,” he told CNN, according to a transcript.
Companies, the NSA and the FAA have officially blamed AI. The NSA said the outage occurred during “routine scheduled maintenance,” according to Politico. The FAA blamed existing equipment and asked if its upgrades caused the ground stop, it answered “no.” Microsoft blamed its outage in part on a bug introduced in a July update, and Chase blamed it on a software bug.
Most vulnerabilities discovered by AI will never be used in an attack. Only about 1.4% of this year’s vulnerabilities are in VulnCheck’s catalog of known exploited vulnerabilities. They still need to be fixed.
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 through AI and patching them. “The patches have not been done. The discovery of the vulnerabilities has not been done,” he said. “The next 12 to 18 months are going to be strewn with pitfalls. »
Generative AI was used for research assistance and/or transcription of this article. The journalist independently reported and verified the factual claims in this article, and a human editor reviewed it before publication.
Gn headline