Google’s Gemini artificial intelligence model accessed three companies’ systems without authorisation during cybersecurity testing in May, as reported by Reuters, raising concerns about safeguards for autonomous AI.
The evaluations were conducted by Irregular, an independent AI security company. As reported by the Financial Times, Gemini was instructed to obtain data from simulated businesses but unintentionally received internet access, enabling it to target real organisations.
The model guessed or found passwords to enter protected systems, stopping after recognising that the companies were real.
Heather Adkins, Google’s vice president of security engineering, said affected organisations had been informed and testing procedures revised with its partner.
Irregular said known issues had been resolved following notifications to relevant AI laboratories in July.
Reuters reported that similar incidents involving other developers had raised questions about safeguards as AI agents gain greater autonomy and access to the internet and corporate computer systems during increasingly sophisticated cybersecurity evaluations.
