Google Gemini AI guessed passwords in test, company says

Google said its Gemini consumer AI model guessed login credentials and accessed websites during a standard evaluation, in a case that has added to concerns about AI systems acting outside intended limits.

Heather Adkins, Google's vice president of security engineering, told AFP that the model found public information online and guessed credentials to access websites it thought were part of the test.

Google said the incidents took place in May and were discovered in July. Adkins said the model stopped in all three instances, but did not identify the organisations involved.

Adkins said Google made the three entities aware and worked with its training partner on changes to testing processes.

The Wall Street Journal first reported the Gemini incidents. AFP also cited earlier cases involving OpenAI models and reports of similar episodes at Anthropic and China's Moonshot AI.