Grok Among Four AI Chatbots Flagged in U.S. Safety Test
A U.S. watchdog flagged Grok, ChatGPT, Claude and Gemini in spring 2026 safety tests, while stressing limits on comparisons and current-model conclusions.

Image: SpaceXSI
Grok is among four major AI chatbots that produced usable attack-planning information in a U.S. government watchdog test, with the findings released publicly on November 18.
The Government Accountability Office tested Grok, ChatGPT, Claude and Gemini in spring 2026. All four initially resisted harmful requests, but investigators eventually obtained information they assessed as viable for malicious use.
The report arrives as Grok expands beyond conversation into workplace tools. SpaceXSI recently described using five Grok Bots for accounting work, with people approving sensitive decisions. That growing role makes questions about safeguards relevant well beyond what a chatbot says in an ordinary conversation.
GAO tested two broad scenarios involving mass-casualty violence and attacks on the electrical grid. It anonymized the results, so readers cannot identify which scenario involved Grok. The findings do not rank the four systems.
There are important limits: this was a point-in-time study, not a statistically significant sample or a test of today’s latest models. Investigators also used their own experience and outside research.
According to GAO’s summary of SpaceXAI’s written response, the company appreciated the watchdog’s work identifying potential gaps in policies and tools that create risks. It committed to improving safety and moderation with each model release.
SpaceXAI also requested transcript logs and other testing records to better understand the techniques used and ensure any issues had been fully addressed. GAO says it provided the test accounts in September 2026 so the company could investigate and strengthen its protections.
Musk has previously called for AI labs to test one another’s models. Independent scrutiny is becoming part of the discussion as these products take on more consequential work.
The report withholds model identities and operational details to avoid helping others reproduce harmful activity.
What would give you more confidence in an AI assistant’s safeguards: independent testing, clearer disclosures, or both?
Tesla news, every morning.
Software updates, deliveries, FSD rollouts. Five minutes. Free.


