Google DeepMind unveils CodeMender, an AI agent that autonomously patches software vulnerabilities – News

Alphabet Inc.’s Google DeepMind lab today shared results for CodeMender, an artificial intelligence-powered agent that automatically detects, patches and rewrites vulnerable code to prevent future exploits.

CodeMember builds on DeepMind’s previous AI-based vulnerability discovery projects such as Big Sleep and OSS-Fuzz, by combining the reasoning power of Gemini Deep Think models with advanced program analysis techniques. The aim is to debug and repair complex security flaws autonomously across massive codebases.

While still only in a research phase, CodeMender has already submitted 72 security fixes to open-source projects, including those spanning more than 4.5 million lines of code. According to DeepMind, CodeMender’s AI-powered agent helps developers and maintainers focus on what they do best — building good software — by automatically creating and applying high-quality security patches.

CodeMender is designed to be both reactive and proactive by instantly patching discovered vulnerabilities and also rewriting existing code to eliminate entire classes of flaws.

In one example, the agent applied “-fbounds-safety” annotations to the libwebp image compression library, the same library exploited in a 2023 zero-click iOS attack. In doing so, it rendered similar buffer overflow vulnerabilities “unexploitable forever,” according to DeepMind researchers.

Under the hood, CodeMender uses a suite of tools including static and dynamic analysis, fuzzing, symbolic reasoning and an “LLM judge” that validates whether proposed changes preserve functionality. The system can self-correct automatically before surfacing its final patch for human review when the validation detects an issue and all changes are verified for correctness, adherence to style guidelines and lack of regressions before submission.

DeepMind notes that CodeMender remains a research effort and that “all patches generated by CodeMender are reviewed by human researchers before they’re submitted upstream.”

The DeepMind team plans to expand outreach to open-source maintainers and “hopes to release CodeMender as a tool that can be used by all software developers to keep their codebases secure,” with technical papers detailing the agent’s architecture and validation pipeline to follow.

If and when it’s released, CodeMember stands in contrast to traditional methods like static analysis and fuzzing that can surface vulnerabilities but still rely heavily on human expertise to validate and repair them. CodeMender’s approach points toward a future where AI systems can handle discovery and remediation, which is arguably a critical step as modern codebases grow exponentially in size and complexity.

Image: News/Ideogram