Google’s AI model Gemini accessed three companies without authorization during online security tests in May. The model stopped the process in all three incidents.
A Google official said Gemini found publicly available information online and, believing it was continuing the test, guessed credentials to access some websites.
Three companies were notified
Heather Adkins, Google’s vice president of security engineering, announced that the three companies were informed of the incidents. Adkins said Google was working with its training partners on changes to the testing processes.
Adkins said the incidents showed that powerful AI models need to be trained to behave responsibly.
OpenAI and Anthropic reported similar incidents
On July 21, OpenAI announced that the AI models used in an internal test to measure cybersecurity capabilities had caused a security breach in the Hugging Face infrastructure.
A few days later, Anthropic announced that its Claude model had gained unauthorized access to the production infrastructure of three different organizations while in a test environment.
U.S. debate focuses on slowing AI development
The rapid spread of AI in the United States has intensified calls to slow development efforts before the technology reaches dangerous levels.
Technology industry representatives, including Anthropic CEO Dario Amodei, OpenAI CEO Sam Altman, and Tesla and SpaceX CEO Elon Musk, have called for AI development efforts to be slowed.
The Trump administration, however, has adopted an approach that encourages innovation rather than slowing AI development. Trump described debates over AI technology getting out of control as a “sick conspiracy” and argued that slowing development efforts would give China an advantage.
