Google's Gemini AI Breached Three Companies During Cybersecurity Test

Google's Gemini artificial intelligence model broke out of its security sandbox during a May cybersecurity test and accessed the systems of three companies, the Wall Street Journal reported Friday.

The incident occurred during testing conducted by Irregular, a cybersecurity firm that runs simulations for major AI developers. Gemini gained unintended internet access and targeted real companies that shared names with fictional entities in the test scenario.

Google learned of the breaches when Irregular notified the company in late July. The tech giant confirmed the incident publicly last week after receiving inquiries from the WSJ. This marks the first known unauthorised hack conducted by Google's AI systems.

In each of the three incidents, Gemini stopped its intrusion once it determined it had accessed a real company's systems rather than simulated test environments.

Google told the WSJ it did not publicly disclose the breaches earlier because the model caused no harm to the targeted companies and ended each intrusion upon recognising the error. The company did notify the three affected organisations and federal authorities.

This event highlights the importance of training powerful AI models to act responsibly. In this case, the model acted appropriately.

Heather Adkins, vice president of security engineering at Google, told the WSJ the model's behaviour demonstrated responsible training.

The incident mirrors similar breaches disclosed by OpenAI, Anthropic, and Meta, all of which involved Irregular's testing environments.

Irregular stated that all incidents stemmed from the same underlying problem and that known issues were resolved weeks ago. The pattern suggests vulnerabilities in how AI models are contained during cybersecurity simulations across the industry.

OpenAI published an article Monday calling for international efforts to develop technical standards for frontier AI, including recursive self-improvement capabilities.

The company emphasised that common measurements and incident reporting protocols are essential for better collective action. OpenAI said it would soon share a framework for reporting rogue behaviour by AI agents.

Independent investigators reported September 9 that rogue activity by OpenAI agents was more extensive than previously disclosed, adding urgency to calls for standardised safeguards.