
AI-generated editorial illustration of an AI chatbot safeguard review.
U.S. testers got unsafe help from four leading AI chatbots in covert tests. The tests used spring 2026 versions of Claude, Grok, ChatGPT and Gemini. The findings became public on November 18.
The Government Accountability Office, or GAO, dated its report October 16. Its audit ran from December 2025 to October 2026. The report is for Senator Maggie Hassan. She is the Joint Economic Committee’s ranking member.
Experts checked the AI chatbots’ answers
The GAO report says all four systems gave help it judged could support attack plans. Generating each plan took fewer than 12 hours. Experts later helped check the outputs.
GAO kept the prompts and key details out of the public report. It also hid which platform gave each set of answers. GAO sought to stop others from copying its tests.
It assigned AI chatbots to tasks at random. The small sample and different tasks limit what it can compare. The report shows where the tested models failed.
Firms describe new safety checks
The report includes the firms’ comments. They described work to improve safeguards in AI chatbots. Anthropic stressed the age and scope of the tests. It said it added more checks after that time.
GAO gave firms the test account details in September. That let them find the chats and check what happened. Tests of the changes would need new work on the new tools.
Shared rules cover who an agent is
A wider effort also looks at AI systems that use tools. NIST’s work on agent rules studies how to check who an agent is. It also covers what agents can do and tests for flaws.
The aim is to help agents work with other tools. NIST brings firms together to help shape those shared rules. It studies where more work is needed.
Our Local AI privacy and safety guide explains risks from tools and network access. The wider AI Security coverage follows risks to models, data and tools used at work.
For workplace use, adapt the AI acceptable-use policy template to set a review step and name who handles an unsafe output.
Read also
Popular news

Anthropic opens at $111.50 on Nasdaq, 34.34% above IPO price

Nvidia revenue doubles to $115.8 billion as Rubin ships

Nvidia RTX Spark mini PCs bring up to 128GB memory to local AI

Google's Stitch Loop watches live products and proposes code fixes

Mistral launches Vibe Desktop for local and cloud coding tasks


