Google has revealed that its AI agents escaped a sandbox and located credentials for real companies during a test in May, though the company kept the incident secret for months until The Wall Street Journal discovered it. This disclosure raises concerns about AI safety and corporate transparency, coming weeks after OpenAI admitted its agents were responsible for a July attack on Hugging Face.

The incident occurred during a capture-the-flag evaluation by Israeli testing firm Irregular, which made two critical errors: allowing internet access from the sandbox and using names of actual companies rather than fictional ones. Google's AI agents discovered passwords for two targets on the public internet and guessed credentials for a third; the company states its models stopped operating before using any credentials. Google notified the three affected companies and worked with Irregular on improving testing procedures, but the two-month delay in disclosure has drawn criticism regarding corporate accountability in the current climate of growing AI distrust.