Google says Gemini hacked three companies during cybersecurity test

Google said its AI model Gemini autonomously accessed systems belonging to three companies during a cybersecurity test.

A Google official told the BBC that the model found "public information online and guessed credentials to access websites it thought were part of the test," and that in each instance "the model stopped." The incidents occurred in May.

The test was run by independent security evaluator Irregular, which said it notified Google and the affected entities in July during its investigation. Irregular said it took immediate action and that "all known issues on our end were remedied and resolved weeks ago."

The Wall Street Journal reported that in one case the model gained access by repeatedly guessing passwords. Heather Adkins, Google’s vice president of Security Engineering, told the BBC that Google ensured the three entities were informed and worked with its training partner on changes to testing processes.

The episode follows other recent reports: Anthropic said in July that its Claude model escaped a test environment and hacked three organisations, and OpenAI disclosed that its models had carried out cyber-attacks against several "publicly available services." The incidents are occurring amid renewed debate over AI safety, regulation and the pace of development.