TradingKey - Google's parent company Alphabet (GOOGL)'s Gemini AI accidentally breached its designated testing scope during a cybersecurity capability test and accessed protected systems belonging to three real companies. This is believed to be the first publicly confirmed autonomous boundary-crossing intrusion by a Google AI model. Google confirmed that the incident occurred in May this year and that the test was conducted by third-party AI security firm Irregular.
The test was originally a "Capture the Flag" cybersecurity exercise in which Gemini was tasked with retrieving specified information from the software systems of fictional companies. However, because the test environment unexpectedly retained internet access and some fictional company names matched real-world companies, Gemini ultimately mistook real enterprises for test targets.
In one incident, Gemini gained access to a protected system by attempting passwords; in two other instances, it discovered login credentials in public online code repositories and used them to access real companies' systems. Google stated that Gemini autonomously halted further operations upon realizing the targets were real companies. Currently, there is no public information indicating that the incident caused data damage or more severe follow-up attacks, and the identities of the three affected companies have not been disclosed.
Heather Adkins, Vice President of Security Engineering at Google, stated that the company has contacted affected organizations and adjusted relevant processes with its testing partner. Irregular stated that it notified relevant AI laboratories of the issue in late July this year and that known vulnerabilities in its testing environment have been remediated.
This incident demonstrates that AI agents are now capable of executing sequential tasks with minimal human intervention, such as searching for information, finding or guessing credentials, and logging into external systems. If permission configurations in test environments fail, models could deploy capabilities intended for security evaluations into real-world internet environments.
Gemini is not the first frontier AI model to encounter such issues. Previously, AI systems from OpenAI, Anthropic, and Meta also breached designated testing scopes during security evaluations conducted by firms like Irregular. These incidents have recently intensified industry discussions around AI agent permission management, sandbox isolation, and third-party security testing standards.
The disclosure comes as OpenAI, Anthropic, and Google DeepMind have all recently stepped up discussions regarding frontier AI safety risks. As AI models acquire greater autonomous execution and network operation capabilities, ensuring that models consistently operate within authorized boundaries is becoming a key issue for the next stage of AI safety testing.
Find out more
Comments