Google DeepMind Introduces CodeMender: AI Agents Who Can Fix Their Own Code Vulnerabilities

JAKARTA - Google DeepMind introduced CodeMender, a new artificial intelligence (AI)-powered agency designed to improve software security by improving code vulnerabilities automatically.

The project was announced on Monday, October 6, 2025 by DeepMind researchers, including Raluca Ada Popa and Four Flynn, as a preliminary result of their latest AI security research.

CodeMender is here to answer the big challenges in the world of software development, which is the length of time it takes to find and patch security loopholes, even with automated methods such as buzzing.

Different from the traditional approach, CodeMender uses the Gemini Deep Think model to analyze, debug, and fix complex vulnerabilities independently.

"During the last six months of the development of CodeMender, we have streamed 72 security improvements to open source projects, including some of which are up to 4.5 million rows of code," the two researchers wrote.

By automatically creating and implementing a high-quality security patch, AI-powered agency CodeMender helps developers and managers focus on important things, namely building good software.

CodeMender is also equipped with other advanced features, such as:

- Root analysis of the cause: Using Gemini, including buzzing and proof of theorem, to precisely identify the fundamental causes of a vulnerability, not just its surface symptoms.

- Self-validated patching: CodeMender is able to provide easy improvements to address complex problems. The agent is not only able to find out the root cause of vulnerability, but is also capable of modifying the system automatically generating and implementing effective patch codes.