cd /news/ai-safety/us-watchdog-got-four-popular-ai-chat… · home › topics › ai-safety › article
[ARTICLE · art-171827] src=ibtimes.co.uk ↗ pub= topic=ai-safety verified=true sentiment=↓ negative

US Watchdog Got Four Popular AI Chatbots To Help Plan Terror Attacks Despite Initial Safety Refusals

A US Government Accountability Office (GAO) watchdog report released on Wednesday, 18 November found that all four tested AI chatbots — ChatGPT, Claude, Gemini and Grok — eventually supplied information that could assist planning of simulated mass-casualty attacks and electricity-grid disruption, despite initially declining some requests, with investigators obtaining the material within 12 hours of testing each platform. The GAO tested the chatbots in spring 2026 at the request of Democratic Sen. Maggie Hassan of New Hampshire, concealed which platform produced each result, withheld the prompts used to overcome safeguards, and described the resulting plans as 'largely viable' while noting flaws such as a proposed attack date that did not coincide with a scheduled event and inaccurate maps. OpenAI said it replayed requests underlying 325 reviewed responses on its newer GPT-6 Astra model, which generally produced more refusals and safer alternatives in an assessment the report did not independently verify, while Anthropic, Google and SpaceXAI said they had strengthened safeguards since the tests.

by read3 min views2 publishedNov 18, 2026
US Watchdog Got Four Popular AI Chatbots To Help Plan Terror Attacks Despite Initial Safety Refusals
Image: Ibtimes (auto-discovered)

Four AI chatbots supplied attack-planning information in US tests, prompting new AI safety concerns #

US government investigators obtained information that could help plan hypothetical violent attacks in American cities from four popular AI chatbots despite initial safety refusals, according to a watchdog report released on Wednesday, 18 November.

The Government Accountability Office (GAO) tested ChatGPT, Claude, Gemini and Grok in spring 2026 after Democratic Sen. Maggie Hassan of New Hampshire requested an investigation into the potential misuse of commercially available artificial intelligence.

How AI Chatbot Safeguards Were Bypassed #

Investigators assigned each chatbot one of two simulated scenarios: an attack intended to cause mass casualties or disruption to the electricity grid designed to inflict economic damage. Two platforms were tested against each scenario.

All four eventually supplied information that the GAO said could assist attack planning, despite initially declining some requests. Investigators obtained the material within 12 hours of testing each platform.

The GAO concealed which chatbot produced each set of results and withheld the prompts and techniques used to overcome safeguards, citing concerns that disclosure could help others reproduce the experiments.

Auditors assessed the information using independent research, expert consultations and site visits, describing the resulting plans as 'largely viable.' But the outputs were not flawless. One proposed attack date did not coincide with a scheduled event, while other responses contained inaccurate maps or overlooked physical barriers.

The agency stressed that it assessed whether information could assist planning, not whether an actual attack would succeed. Some verification steps were modified for safety, and not every instruction was tested.

The exercise was also limited to the versions available at the time. It was not statistically representative, and because the platforms were assigned different scenarios, their performance cannot be directly compared.

AI Chatbot Companies Challenge the Findings #

Anthropic questioned whether the testing showed that AI offered an advantage over ordinary internet searches. The GAO acknowledged that a controlled comparison would be useful but said measuring any speed advantage was outside the investigation's scope.

OpenAI argued that the report did not consistently distinguish ChatGPT's contributions from information supplied by investigators, outside research and the GAO's own analysis. The agency maintained that its purpose was to examine how commercially available AI could be used maliciously.

OpenAI said it subsequently replayed requests underlying 325 reviewed responses using its newer GPT-6 Astra model, which generally produced more refusals and safer alternatives. That assessment was conducted by the company, not independently verified in the report.

Google said the tests involved earlier models and pointed to ongoing improvements in its safeguards. SpaceXAI, which operates Grok, also said it had strengthened safety measures since testing. Anthropic likewise described additional protections introduced after the tests.

The report also examined companies' reactions after suspicious exchanges. One suspended an investigator's account after testing ended. Two others removed portions of chat histories, although auditors could not determine whether the removals reflected safety measures or errors.

Hassan said she would make AI safety an early focus of Senate homeland security oversight and called for stronger safeguards before models are released, regular updates and reporting of illegal activity to law enforcement. She warned that without congressional action, a major AI-assisted attack could become 'a question of when, not if.'

© Copyright IBTimes 2026. All rights reserved.

── more in #ai-safety 4 stories · sorted by recency
── more on @government accountability office 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/us-watchdog-got-four…] indexed:0 read:3min 2026-11-18 · —