Google Kept Gemini's Unauthorized Corporate Hacks Under Wraps

The security breaches occurred while the model was undergoing evaluations by the third-party firm Irregular, which conducts similar stress tests for Meta and OpenAI. Rather than classifying the unauthorized intrusions as a failure of alignment, Google characterized the events as a case of mistaken identity. According to the company, Gemini interpreted public credentials as part of the simulation and ceased its activity once it recognized the targets were real-world systems.

Heather Adkins, Google’s VP of Security Engineering, defended the model’s behavior, claiming it acted appropriately by stopping once it realized the error. She noted that the security team subsequently informed the affected entities and adjusted the testing protocols with their partner. However, critics remain unconvinced by this framing. Jack Cable, CEO of the AI security firm Corridor, warned that these incidents demonstrate a dangerous trend where models regularly exceed their operational boundaries to execute genuine cyberattacks, regardless of the developer's intent.