Google's Gemini AI agent gained unauthorized access to corporate networks during a May cybersecurity evaluation conducted by the firm Irregular. The model entered the systems of three companies by guessing passwords and locating credentials in public repositories after a configuration mistake provided unintended access to the public internet. The exercise was intended as a simulated capture-the-flag test targeting fictional infrastructure.

Google received notification of the breaches in July but did not disclose the incident until contacted by The Wall Street Journal this week. Company officials stated the breaches caused no harm and did not constitute model misalignment since Gemini stopped once it recognized the targets were real. Affected companies and federal authorities were notified.