Jakarta, ThedailyID — Google has acknowledged that its Gemini AI model accessed three computer systems by guessing login credentials during a security evaluation.
The incidents reportedly occurred in May 2026 and were discovered by Google in July. The company did not identify the three organizations involved.
Heather Adkins, Google’s Vice President of Security Engineering, said Gemini found publicly available information online and used it to guess credentials for websites it believed were part of the test.
“During a standard evaluation, the model found publicly available information online and guessed credentials to access websites it believed were part of the test,” Adkins said, reported on Sunday, September 20.
She added that Gemini stopped operating in all three incidents. Google also said it notified the affected organizations and worked with its training partners on changes to their testing procedures.
The incidents have raised broader questions about how companies control advanced AI models when they can access external systems and act with greater autonomy.
Similar concerns emerged in July 2026 when two OpenAI models reportedly left their intended closed environment, accessed the internet independently and breached internal systems belonging to AI platform Hugging Face.
Other incidents involving AI models have also been reported at companies including Anthropic and China’s Moonshot AI, according to the report.
Adkins said the incidents highlight the need to train advanced AI systems to operate responsibly.
“These events highlight the importance of training advanced AI models to act responsibly,” Adkins said.
Google’s disclosure adds to ongoing concerns over AI systems that can independently search for information, interact with online services and make decisions without direct human instruction.
