During a test conducted by the Israeli cybersecurity company Irregular to assess advanced artificial intelligence models, Google’s Gemini model reached the electronic infrastructure of three real companies because of unplanned internet access.
The test was intended to take place in a closed environment, with Gemini carrying out cyberattacks only against test systems. However, the accidental activation of the internet connection allowed the model to move beyond the systems included in the test.
Gemini found or guessed login credentials
Google said Gemini used information it found online or guessed login credentials to connect to websites it believed were part of the test. The model was stopped in all three incidents.
The company said the three companies affected by the incidents had been informed, but that the incidents had not been disclosed publicly. Google also said no damage had occurred and that a public announcement was not considered necessary.
Heather Adkins, Google’s vice president responsible for security engineering, said the incidents showed that powerful artificial intelligence models needed to be trained to operate responsibly.
Irregular said the problems had been resolved
Irregular said the incident was linked to the same case reported previously and that, in essence, no separate incident had occurred. The company said the relevant research laboratories and affected parties had been informed at the end of July 2026, while the identified problems had been resolved weeks earlier.
Irregular had previously conducted security tests for OpenAI and Anthropic. The company had also encountered separate incidents in which artificial intelligence models moved beyond their testing environments and reached real systems.
US debates oversight of development efforts
The incident during the Gemini test reignited debate in the US over the oversight of advanced artificial intelligence models. Senator Bernie Sanders called on companies to temporarily halt the development of advanced models, saying what happened demonstrated how difficult it was to control the technology.
OpenAI announced that it had halted advanced model development efforts for two weeks. Dario Amodei, Anthropic’s chief executive officer, called for artificial intelligence work to be slowed collectively and systematically to ensure that the necessary safety mechanisms were developed.
