Google Gemini learned to find and exploit vulnerabilities in third-party software

Company Google reported on the testing results of its Gemini model, which demonstrated the ability to autonomously search for and exploit vulnerabilities in third-party software developers' products. During controlled tests, the neural network successfully detected security breaches and conducted attacks on target systems.

According to the report, the model was configured to perform cybersecurity tasks as part of an internal testing program. In the process, Gemini independently identified attack vectors by analyzing source code and service configurations. The developers emphasize that the model acted strictly within the set parameters, completing each action immediately after confirming the vulnerability.

The company stated that the model "acted adequately" and did not cause real damage to the infrastructure. The use of such tools is aimed at automating the bug bounty process and increasing the security of corporate systems before they can be exploited by attackers. All actions of the neural network were recorded in logs for subsequent security audits.

This incident confirms the trend of integrating autonomous agents into cybersecurity processes. Earlier, other models demonstrated similar capabilities, but the current Google case stands out for the scale of testing on real external objects. Developers continue to improve the mechanisms of the "ethical limiter" to prevent the unauthorized use of the model for destructive purposes.