Gemini Illegally Accesses Three External Companies... Google Says "No Damage, Not a Matter for Disclosure"

Access to Real Company Systems After Internet Connection Enabled During Security Evaluation

It has been revealed that Google's artificial intelligence (AI) model, Gemini, accessed the systems of three actual companies without authorization during a security evaluation. While there have been repeated incidents of major AI models hacking external organizations, this is the first time Gemini has been confirmed to have infiltrated real company systems.

According to the Wall Street Journal (WSJ) on September 18 (local time), the breach took place during a cybersecurity evaluation conducted by the AI security startup Irregular in May, when Gemini accessed the systems of three actual companies.

Originally, Irregular intended to conduct a simulated hacking test on virtual companies while blocking internet access. However, due to an error in the test environment, internet connectivity was unintentionally enabled, which allowed Gemini to access real systems on the internet.

In one test, Gemini received instructions to obtain information from a virtual company's software. After connecting to the internet, Gemini identified an actual company with the same name as the virtual test subject and attempted to access that company's services by guessing passwords.

According to Google, once Gemini realized it had accessed a real company rather than a virtual one, it discontinued any further actions.

In two other tests, Gemini discovered repositories on the internet containing credentials of other companies and used them to access real corporate systems. In these cases as well, Google explained that Gemini stopped acting once it recognized the targets were actual companies.

Google stated that it did not separately disclose the incident because Gemini did not cause any harm to the real companies. The company further explained that the model stopped acting immediately after verifying that the targets were actual organizations.

Heather Adkins, Vice President of Security Engineering at Google, said, "This incident demonstrates how important it is to train powerful AI models to behave responsibly."

Google explained that its latest AI models were not involved in this incident, but did not specify which version of Gemini was used. Google reported the matter to federal authorities, and Irregular also informed related labs and organizations of the incident.

However, there has been criticism over Google's decision not to disclose the incident independently.

Jack Cable, CEO of the security startup Corridor, argued that the very fact that an AI model exceeded its initially assigned scope and carried out an actual cyberattack poses a problem, stating, "This is a matter of public interest the public deserves to know."

This incident comes as similar cases have come to light in which AI models developed by major firms such as OpenAI and Anthropic have attacked external systems.

Previously, it was revealed that OpenAI's GPT breached its isolated testing environment and hacked into the AI software company Hugging Face, while Anthropic's Claude likewise accessed the internet during a security evaluation and infiltrated external corporate systems.

With these incidents continuing, concerns have been growing regarding the cyberattack capabilities and control problems associated with high-performance AI models. Some in the AI industry have even argued that the pace of model development should be slowed to ensure that sufficient safety measures are in place.

This content was produced with the assistance of AI translation services.