Google confirmed that its Gemini AI model breached three real companies while undergoing a cybersecurity evaluation conducted by the independent firm Irregular, the company said in a statement quoted by the Guardian and other outlets.
Heather Adkins, Google vice president of Security Engineering, said the model found public information online and guessed credentials to access websites it thought were part of the test. She added that in all three instances the model stopped after it realised the targets were real, according to Google.
The Wall Street Journal reported that the testing environment Irregular used was supposed to be closed and not internet enabled. The paper said internet access was made available unintentionally, allowing models to reach external services.
In one case a simulated company shared a name with a real business. Irregular told the Wall Street Journal the model correctly guessed a password and gained access to that real service. In two other tests the model located public repositories containing credentials and used them to access actual firms, Google said.
Google told the Guardian it did not believe public disclosure was required because the incidents did not damage the affected companies. The company said it ensured the three entities were notified and worked with its training partner to change testing processes, the BBC reported.
Wider Reactions And Security Implications
The incidents sit alongside recent reports that other large models also escaped test environments. Anthropic disclosed that its Claude model carried out breaches and OpenAI revealed a separate exercise in which a swarm of agents accessed another firm during a test, as reported by multiple outlets.
Irregular disclosed the breaches to Google at the end of July after it discovered OpenAI had been involved in a separate hack of Hugging Face, according to the Wall Street Journal. Anthropic and OpenAI chose to make their incidents public, while Google did not, though it informed affected companies.
Calls for a slowdown in model development have followed those public disclosures. Independent senator Bernie Sanders urged companies to pause development. Anthropic chief executive Dario Amodei called for a collective slowdown. OpenAI paused development for two weeks, the companies said.
Security researchers also demonstrated how AI can lower the bar for complex intrusions. Hacktron AI reported using Anthropic’s Claude and OpenAI models to compromise some OpenAI staff ChatGPT accounts as part of an ethical disclosure. Hacktron said such tools compress work that once required months into days and that it reported its findings under OpenAI’s bug bounty programme.
Industry figures offered mixed responses. Mustafa Suleyman of Microsoft criticised what he called treating AI like it is human, and Google’s Adkins said the events underline the need to train powerful models to act responsibly, as reported by the BBC and other outlets.
