Google CodeMender Is Now in Preview - an AI Agent That Finds, Proves, and Fixes Code Vulnerabilities

Google launched CodeMender in preview on Monday, an AI security agent that scans codebases for vulnerabilities, confirms whether they are actually exploitable, and generates patches - all without handing the manual verification queue back to a developer. The agent runs via Gemini Enterprise Agent Platform today, or as part of Google's AI Threat Defense product alongside Wiz. Gemini 3.5 Flash Cyber, announced the same day as a model purpose-built for vulnerability work, will power a separate government-only access tier.

Scan, Verify, Remediate - the Three Steps CodeMender Handles

Scanning covers the vulnerability classes developers actually lose sleep over: memory corruption, injection flaws, web security issues, cryptographic problems, and insecure data handling. CodeMender supports C/C++, Go, Java, Python, Ruby, Rust, and TypeScript. Unlike static analysis tools that flag patterns and leave triage to humans, CodeMender uses contextual understanding of what a specific repository is trying to do - which cuts down on findings that are technically a match but pose no real risk.

Verification is where CodeMender does something most scanners skip. Rather than reporting a potential flaw and moving on, the agent builds a proof-of-concept exploit and runs it in an isolated customer-managed sandbox. Only vulnerabilities that survive that test - confirmed as exploitable - move forward for patching. The result is a prioritized list where every item has already been proven dangerous, not just suspicious.

Remediation closes the loop. CodeMender generates a secure patch, delivers it as a code diff inside VS Code or Google's Antigravity IDE, and then uses an LLM-as-a-judge step to verify the fix does not break existing application logic. Developers manually review and approve before anything is committed. Source code is never retained - Google routes traffic through VPC and keeps data isolated per enterprise security requirements.

Two Access Paths, One Government-Only Tier

For most enterprise customers, CodeMender is accessible through Gemini Enterprise Agent Platform or as part of AI Threat Defense. The AI Threat Defense path adds Wiz to the picture: Wiz orchestrates the full application security lifecycle, calls CodeMender to scan code, enriches findings with deployment context from the Wiz Security Graph, and triggers Wiz Red Agent for AI-based pentesting. The system flags highest-risk vulnerabilities first, narrowing what developers actually need to look at.

A third tier exists for governments and trusted partners only: CodeMender powered by Gemini 3.5 Flash Cyber, a model trained specifically for security research. That access path has no general availability date. Fewer than 5% of AI-identified open source vulnerabilities currently get patched - which is the actual problem CodeMender is trying to close, not the scanning part that most tools already handle adequately.

Multi-Model and Third-Party Support Coming

Google designed CodeMender for flexible model selection - teams can optimize for cost, scanning speed, or depth depending on the job. Third-party frontier model support is coming later in 2026, meaning CodeMender won't be locked to Gemini once that ships. Salesforce, Robinhood, and Palo Alto Networks are named early customers in Google's launch post, each citing the automated path from finding to tested fix as the capability that distinguishes it from existing tools.