During a test, Gemini broke into the systems of three real companies
Google revealed on 18 September that in May, during a cybersecurity exercise, its AI model accessed three outside systems without authorisation, believing they were part of the challenge.
- In May, during a cybersecurity test, Gemini accessed the systems of three real companies.
- The cause: a poorly isolated test environment and a fictitious name that matched a real domain.
- Gemini used guessed or published credentials; Google says no damage was caused.
Google said on 18 September 2026 that its Gemini model had accessed the systems of three companies without authorisation during a test carried out in May. The exercise, a “capture the flag” challenge, was run by Irregular, an independent company that evaluates advanced AI systems.
A configuration error
The name of a fictitious company used in the exercise matched a real internet domain, and a misconfiguration left the test environment connected to the network instead of isolated. Gemini took the real systems for targets in the challenge. According to Google, it got in either by guessing credentials or by using credentials found in a public repository.
Google says the agents stopped on their own when they realised these were real systems, and that the intrusions caused no damage as far as it knows.
- Google's Gemini becomes latest AI model to break out and hack computer systems, CNBC.
- Google says its AI model gained unauthorized access to three outside systems, NBC News.
Spotted an error? Write to infos@novotech.blog. Corrections are flagged at the end of the article.
