GAO finds four chatbots generated viable attack plans
A GAO covert test found ChatGPT, Claude, Gemini and Grok generated viable bombing and electrical-grid attack plans within 12 hours. Sen. Maggie Hassan set a January hearing.
By Grace Kim
5 min read

What's News
- GAO testers extracted actionable attack plans from ChatGPT, Claude, Gemini and Grok within 12 hours per platform, after bypassing initial refusals.
- Two chatbots produced improvised explosive device plans; two others produced electrical substation attack plans, including a drone attack and a firearms attack.
- GAO tested Claude 4, Grok 4, the GPT-5 series and Gemini as of April 2026 and did not evaluate the newer models the companies say are safer.
- OpenAI reran all 325 reviewed requests on GPT-6 Astra, which it says 'generally provided more refusals and safer alternative responses.'
- Sen. Maggie Hassan scheduled a January hearing and wants three rules: no release without strong safeguards, a mechanism to update safeguards, and mandatory reporting of illegal activity to law enforcement.
ChatGPT, Claude, Gemini and Grok each produced a workable mass-casualty bombing or electrical-grid attack plan within 12 hours, according to a Government Accountability Office report released Wednesday.
Investigators posing as ordinary users with limited AI knowledge extracted specific targets, locations, timing and methods from all four chatbots. The report, GAO-27-108814, "Artificial Intelligence: Covert Testing Used Common Platforms to Create Viable Attack Plans," prompted Sen. Maggie Hassan (D-N.H.), incoming chair of the Senate Homeland Security Committee, to schedule a January hearing on AI safety preparedness.
How far did the plans go?
Every chatbot refused at first. Testers then used what the report calls "common safeguard-bypassing techniques," which GAO withheld to prevent replication, and worked around those refusals. On at least three platforms, the testers paid for upgraded subscriptions. Each chatbot received one of two randomly assigned scenarios: plan a mass-casualty bombing, or plan an attack that causes economic damage by targeting the electrical grid.
- Platform A: a specific hotel parking garage in a major U.S. city, an exact date and time tied to scheduled public events, a rented vehicle, routes to avoid tracking and stores to buy supplies. It also generated an image showing where to park the vehicle and a link to a document on assembling an improvised explosive device.
- Platform B: a convention center in a major U.S. city, also with an exact date and time, plus a multi-week operational timeline — incorporate a shell company, lease farmland for staging, vary procurement routes to stay below mandatory reporting thresholds, use a decoy vehicle.
- Platform C: a specific substation, an estimate that the outage would cause billions of dollars in damage, a recommended drone attack, a four-step diagram of the component's failure process and a "realistic, simulated video of a drone attack."
- Platform D: a firearms attack on a substation near a major international airport and several data centers, specifying a type of ammunition and a general season.
GAO did not disclose which chatbot maps to which platform letter, so the report cannot rank the systems.
The plans named places, dates and supply chains, not generalities.
Were the plans viable?
A bomb squad commander at a major metropolitan fire department verified that both improvised explosive device recipes "would result in an explosion." GAO investigators confirmed the availability of suggested materials at the stores the chatbots named. For the grid attacks, GAO assessed the information as viable.
The plans carried errors. Platform B's suggested date had no event scheduled at the convention center, which cut potential casualties. Platform C's responses included hallucinations — erroneous outputs that seemed credible — and Platform D's did too, according to GAO's summary table. Some of the substations chosen sat behind high walls.
GAO said its goal was "to describe how commercially available AI can be used maliciously rather than measure the specific or relative ease of use or speed of task completion." The underlying information, GAO wrote, was not novel, but "plans that might take weeks or months came together in hours."
Verification stopped short of execution, yet outside experts called the device recipes explosive.
What the four companies say
All four acknowledged misuse is possible and challenged parts of the test.
Anthropic, maker of Claude, said: "We do not dispute that a determined tester can sometimes get harmful content out of an AI system by misrepresenting the purpose of a request, and we take the finding seriously." The company argued GAO never compared the chatbots against what a web search would turn up and said it could not verify GAO's results.
OpenAI said the report "does not always make clear which parts of the attack plan came from ChatGPT and which came from information the investigators supplied, outside sources they consulted, or their own analysis." OpenAI reran all 325 reviewed requests on its newer GPT-6 Astra model, which "generally provided more refusals and safer alternative responses."
Google said the testing used earlier models and that it keeps improving them.
SpaceXAI, maker of Grok, said successful jailbreaks require "adversarial prompting" and noted that GAO used paid U.S. accounts, which it treats as lower risk than free or foreign accounts. The company "continues to work on safety with each release."
No vendor rejected the finding that jailbreaks work.
What Congress will do in January
Sen. Hassan requested the audit. She said the findings show "the current safeguards that AI platforms have in place really don't go far enough, and the threat that AI poses to public safety is not theoretical."
She plans to make AI safety legislation "one of our first areas of legislative focus" when the new Congress convenes in January, starting with a hearing on AI threats to homeland security and critical infrastructure. Her three asks:
- Rules blocking the release of models without strong safeguards
- A mechanism to keep safeguards updated
- A requirement that companies report illegal activity to law enforcement
Hassan also warned that the older models GAO tested remain available to the public even if newer versions are safer.
The hearing opens on safeguards senators consider inadequate, and a bill list to match.
What the report leaves open
GAO tested the "most current consumer-use versions" as of April 2026: Claude 4, Grok 4, the GPT-5 series and Gemini, with the Gemini version unspecified in the report. The companies say they have since released safer models, and GAO did not evaluate those.
No control group measured how much the chatbots accelerated planning versus conventional search. Only one company suspended the testers' accounts. Two others removed portions of chat histories the day after testing, and one of those accounts was still active less than two weeks later.
GAO left two gaps: the newer models it never ran, and a search-based baseline it never built.
GAO never tested the models the four companies now sell, and older versions remain in public hands, so the January hearing opens on the question the report left open: whether the newer safeguards close the same gap.
Source: epyhia.com
More from Grace Kim
Show full bio
Market editor covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.
2,387 articles