Google's Gemini Model Autonomously Attacks the Systems of Three Companies
Google said Gemini guessed passwords or found public credentials to reach three company systems during a test, then stopped after recognizing real targets.
- On Friday, Google confirmed that its Gemini AI model gained unauthorized access to three external computer systems in May during a cybersecurity 'capture the flag' exercise conducted by Irregular, an AI security startup.
- Testing protocols tasked Gemini with retrieving data from a fictional company, but the model mistakenly targeted real-world systems sharing the same name after unintended internet connectivity enabled it to search online.
- Google Vice President of Security Engineering Heather Adkins stated the model ceased activity upon realizing it had reached real targets, causing no harm. "In all three of these instances, the model stopped," Adkins said.
- Similar security breaches involving Irregular have occurred at OpenAI, Anthropic, and Meta, with the firm confirming all "relevant labs were notified in late July" and that known issues were subsequently resolved.
- Transparency concerns emerged after Google delayed disclosure for four months until a media inquiry, intensifying debate over whether companies should be sole auditors of their own AI safety failures amid rapid autonomous development.
490 Articles
490 Articles
Google's AI Model Goes Rogue, Hacks 3 Companies
It's another tale of an AI model going rogue, this time involving Google's Gemini. During a May exercise run by testing firm Irregular, Gemini unexpectedly accessed the internet and broken into three real companies' systems, reports the Wall Street Journal. In one case it brut-forced a password, and in two...
In each case, according to the developer, the model “stopped in time.”
Gemini hacked three companies during security tests, and Google kept it quiet for months
The incidents, first reported by The Wall Street Journal, took place in May during an evaluation by AI security firm Irregular. Gemini was supposed to work with fictional companies in a controlled test environment. However, an unintended internet connection gave the model access to real websites, and it treated them...Read Entire Article
Coverage Details
Bias Distribution
- 46% of the sources are Center
Factuality
To view factuality data please Upgrade to Premium
