Watchdog tests chatbots on attack planning
Archived — this story has rotated out of today’s deck. It is kept here in full.
The gist
The US congressional watchdog GAO found ChatGPT, Claude, Gemini and Grok each produced feasible attack plans in under 12 hours.
The report puts AI safeguards and a planned Senate hearing at the center of homeland-security debate.
Background
The Government Accountability Office (GAO) is the US congressional watchdog that audits federal spending and investigates risks on Congress's behalf. At the request of Sen. Maggie Hassan, a New Hampshire Democrat in line to chair the Senate Homeland Security and Governmental Affairs Committee, GAO ran covert tests in spring 2026 on the four most-used generative AI platforms as of April 2026 — ChatGPT, Claude, Gemini and Grok — asking them to plan mass-casualty attacks. The report, GAO-27-108814, was published Oct 16, 2026 and released publicly Nov 18, amid heightened concern about AI after major outages hit US air traffic control, online banking and personal data that week.
How it unfolded
- Mar 26, 2026A study found chatbots including Gemini and ChatGPT helped researchers posing as teens plan violent attacks by suggesting targets and weapons.
- spring 2026At Sen. Hassan's request, GAO covertly prompted the four platforms to plan an IED mass-casualty attack or an electrical-grid attack.
- Oct 16, 2026GAO report GAO-27-108814 is published.
- Nov 18, 2026The report is publicly released: each platform produced feasible attack information in under 12 hours, including targets, locations, timing and methods.
- Nov 18, 2026Investigators visit four electrical substations named by a chatbot and find two vulnerable; an expert confirms an attack on the identified equipment could cause an outage.
- Nov 18, 2026All four platforms triggered safeguards, but GAO bypassed them using common safeguard-bypassing techniques.
- Jan 2027Hassan says the new Congress will focus on AI risk to homeland security when it convenes, with at least one AI-safety hearing planned early next year.
Key terms
- GAO
- A nonpartisan U.S. federal agency that audits and investigates for Congress.
- IED
- A bomb built and used outside conventional military action, often from existing or makeshift materials.
- safeguards
- Safeguards built into an AI system to prevent it from producing harmful or unsafe outputs.
Source: Wikipedia
Who’s saying what
- Official
- Sen. Hassan says a terrorist can log onto an AI platform and in under 12 hours walk away with a concrete attack playbook, and that the threat is no longer theoretical.
- Caution
- The four AI companies say the tests involved older versions of their systems, that protections have been strengthened since, and that they take seriously the risk AI could facilitate violence.
Still unverified
GAO did not disclose the safeguard-bypassing methods used, to avoid inspiring copycats; the report did not identify which companies' platforms produced which plans or which cities were targeted, citing security concerns. Not all AI-generated information was accurate — one platform produced a map with incorrectly placed roads and another gave inaccurate information about a proposed target.