Master CodeMender with our step-by-step tutorial, detailed feature walkthrough, and expert tips.
Explore the key features that make CodeMender powerful for voice agents workflows.
CodeMender is an AI agent for code security developed by Google DeepMind, announced in late 2025. It uses Gemini Deep Think reasoning models combined with program analysis tools to autonomously identify, patch, and rewrite vulnerable code. The project is part of DeepMind's broader AI safety and responsibility initiative. It has already contributed 72 security fixes to open-source codebases.
As of its late 2025 announcement, CodeMender is not publicly available — there is no signup page, API, or self-serve product. DeepMind is gradually reaching out to maintainers of critical open-source projects to upstream patches collaboratively. The team has stated they plan to release technical papers and engage with the security research community over time. For most developers, the practical path today is to monitor DeepMind's blog and security-focused publications for updates.
Unlike Copilot Autofix or Snyk DeepCode, which primarily suggest fixes for developers to review, CodeMender autonomously generates, validates, and self-critiques patches using fuzzing, SMT solvers, and differential testing before any human review. It also goes proactive — rewriting code with hardened APIs and compiler annotations like -fbounds-safety to eliminate entire vulnerability classes rather than fixing one bug at a time. Based on our analysis of 870+ AI tools, this combination of autonomous patching plus formal validation is rare in the category.
CodeMender targets a broad range of software vulnerabilities, with public demonstrations focusing on memory safety issues such as buffer overflows in C/C++ code. Its work on libwebp showed it can apply -fbounds-safety annotations that would have prevented the CVE-2023-4863 zero-click iOS exploit and many similar buffer-overflow vulnerabilities. The agent uses root-cause analysis rather than surface patching, meaning it addresses underlying logical flaws rather than just visible symptoms. DeepMind has indicated broader language and vulnerability-class coverage is part of ongoing research.
Every patch goes through a multi-stage validation pipeline before human review. CodeMender runs the modified code against existing regression test suites, executes fuzzers to catch runtime issues, and uses differential testing to compare behavior before and after the change. An LLM-based self-critique agent then reviews the patch for correctness, regressions, and quality issues. Only patches that pass all automated checks are surfaced for human security researchers to review and upstream.
Now that you know how to use CodeMender, it's time to put this knowledge into practice.
Sign up and follow the tutorial steps
Check pros, cons, and user feedback
See how it stacks against alternatives
Follow our tutorial and master this powerful voice agents tool in minutes.
Tutorial updated March 2026