The United States Government Accountability Office (GAO), the investigative arm of the US Congress, on Wednesday, 18 November 2026, made public a report in Washington saying its undercover investigators got the four most widely used generative AI platforms, OpenAI's ChatGPT, Anthropic's Claude, Google's Gemini and Grok, to produce plans that could help someone carry out an attack inside the United States. Each platform did so in less than 12 hours, despite safeguards meant to stop exactly this kind of request.

The 61-page report was requested by Senator Maggie Hassan, a Democrat from New Hampshire who is due to chair the Senate Homeland Security committee when the new Congress meets in January 2027. She has said she will hold hearings on AI safety early next year. Here is what the report says, how the companies have responded, and what remains unclear.

What did the GAO actually find?

The GAO found that all four platforms gave out information it judged "feasible" enough to help a malicious actor plan and execute an attack. According to the report's public summary, titled "Artificial Intelligence: Covert Testing Used Common Platforms to Create Viable Attack Plans" (GAO-27-108814), two platforms produced material on an explosive-device attack aimed at causing mass casualties, while the other two produced material on attacks against the electrical grid meant to cause economic harm.

The platforms' safeguards did kick in. The GAO says its requests were at times refused, content was removed and accounts were suspended. But the watchdog says it got around those protections using "common" techniques for bypassing safeguards. The Associated Press, which reviewed the full report, reported that one platform at first declined to give detailed instructions on an explosive device before investigators manipulated it into doing so.

According to the AP and CBS News, the plans included possible targets, locations, timing and methods, and some came with maps and images. The GAO has withheld operational detail from the public version and does not say which chatbot produced which plan; it assigned each platform a letter at random when presenting results.

How was the testing done?

  • Audit period: December 2025 to October 2026, with the covert tests run in spring 2026, according to the AP.
  • Choosing the platforms: the GAO says it picked the four most-used platforms based on usage statistics as of April 2026.
  • Choosing the scenarios: working with experts, the GAO settled on two misuse scenarios, a mass-casualty explosive attack and an attack on the power grid, and randomly assigned each platform to one of them.
  • Checking the output: investigators tested parts of the plans through research, surveillance, site visits and interviews with experts. They did not carry out any attack.

The checks also found mistakes. The AP reported that in one case roads were placed wrongly, and in another, high walls would have blocked the view of an intended target.

Timeline: how the report came about

  • December 2025: GAO begins the performance audit requested by Senator Hassan.
  • April 2026: usage figures from this month are used to pick the four platforms.
  • Spring 2026: covert testing takes place.
  • October 2026: audit work ends; the report is dated 16 October 2026, according to the GAO's product page.
  • 18 November 2026: the report is released to the public.
  • January 2027: the new Congress convenes; Hassan plans one of its first hearings on AI risks.

Verified data at a glance

  • Platforms tested: 4 (ChatGPT, Claude, Gemini, Grok).
  • Scenarios: 2 (explosive device for mass casualties; attack on the electrical grid), two platforms each.
  • Time taken: under 12 hours per platform, though the GAO notes this could vary by topic and user.
  • Report length: 61 pages, including the companies' responses, according to the AP.

What the GAO says the AI actually added

An important nuance sits inside the report itself. The GAO says the information the chatbots supplied was largely not new and could be found elsewhere. What the AI did, in the watchdog's view, was pull it together into usable plans far faster than a person could on their own, compressing what CBS News described as weeks or months of planning into hours. "The potential benefits of this technology are immense, but they can also benefit malicious actors," the GAO wrote, as quoted by Fox News and the AP.

What do the companies say?

The four companies' responses are included in the report. According to the AP:

  • Anthropic, which makes Claude, said it does not dispute that a determined tester can extract harmful content by misrepresenting the purpose of a request, and that it takes the finding seriously. It also said it could not verify the GAO's findings, and Fox News reported it questioned aspects of the methodology.
  • Google said the testing was done on earlier versions of its models and that it continues to improve them.
  • OpenAI said it welcomed the chance to respond to the findings and acknowledged the work that remains.
  • SpaceXAI, named as responsible for Grok, said it continues to work to improve safety with each model release.

The GAO says company officials generally told it they would keep improving the safety of their platforms.

What lawmakers say

Senator Hassan said on a conference call, "We need to act now." In a statement reported by the AP and Fox News, she said the public-safety threat from AI is "no longer theoretical" and that, without action by Congress, it is "a question of when – not if" a bad actor uses AI to plan and carry out a major attack. No specific bill has been announced so far.

The report lands in a week of scrutiny over how AI systems behave in sensitive settings. MapMyTimes reported earlier on a Vals AI study in which leading models marked "targets" in drone and satellite images, and on California's lawsuit against OpenAI.

Why does this matter for readers in India?

The answer is that the same consumer chatbots tested by the GAO are used in India, so the weaknesses it describes are not confined to the US. Google, for example, recently included India among six countries for a free rollout of Gemini's Tasks mode. Safety changes the companies make after a report like this are typically applied to the global product, though the GAO report itself covers only US-focused scenarios.

India's own approach so far leans on existing law rather than a new AI statute. When the Ministry of Electronics and Information Technology (MeitY) released the India AI Governance Guidelines on 5 November 2025, Secretary S. Krishnan said the government's focus "remains on using existing legislation wherever possible", while Principal Scientific Adviser Ajay Kumar Sood described "Do No Harm" as the guiding principle. The US debate that Hassan now plans to lead is likely to be watched by Indian policymakers and the many Indian businesses building on these platforms.

What is established and what is uncertain?

Established

  • The GAO report GAO-27-108814 exists, was publicly released on 18 November 2026, and states that all four tested platforms produced information the GAO judged feasible for planning attacks.
  • Each platform did so in under 12 hours, and safeguards were triggered but bypassed, according to the GAO.
  • None of the companies has publicly denied that safeguards can be bypassed; Anthropic explicitly accepted that a determined tester can extract harmful content.

Uncertain or disputed

  • Which platform produced which plan; the GAO has not said.
  • Whether the same techniques would work on current model versions. Google says earlier models were tested, and the tests took place months ago.
  • Whether the results can be reproduced; Anthropic says it could not verify them.
  • How the platforms compare with one another on safety, since the GAO has not published platform-by-platform results.
  • What, if any, legislation will follow.

What happens next?

The next step is in Congress. Hassan plans to hold one of the first hearings of the new Congress in January 2027 on AI risks and how to improve safety, according to the AP. The companies say they will keep updating their safeguards. What, if anything, Congress does about AI safeguards should become clearer once those hearings begin.

Photo source