Anthropic has added a security-guidance plugin to Claude Code that reviews code for common vulnerabilities while developers work. The company says it can catch issues before pull requests, and that internal use reduced security-related PR comments by 30% to 40%.
Anthropic has introduced a security-guidance plugin for Claude Code that reviews code changes for common vulnerabilities and helps fix them during the same development session. The company says the goal is to catch problems before code reaches pull requests, reducing later manual security review.
The plugin is designed to run automatically during development sessions. According to Anthropic, developers do not need to launch separate tools or remember extra commands once it is installed.
This news matters because it adds security checks directly into the coding assistant, instead of leaving security review for later. A pull request, or PR, is a proposed code change that teammates review before it is merged. By checking for common bugs like injection flaws and unsafe deserialization early, Claude can reduce how many issues reach human reviewers.
If you build with self-hosted agents, add a fast local rules layer that flags dangerous functions and unsafe output handling before the agent writes a commit. Keep a second pass over the full diff so the system can catch cross-file issues like authorization bypass and SSRF, then run a third check only on commit or push to validate surrounding code. Also make your repo-specific rules editable, so teams can add checks for internal sanitizers, approved libraries, and prohibited system calls without retraining the agent.
Anthropic said the plugin focuses on issues such as injection flaws, unsafe deserialization, and insecure DOM APIs. It is intended to act as a first pass that flags problems early, while code is still being written.
The system works in three review stages. The first stage runs during file edits and uses lightweight pattern checks without calling a model. These checks look for risky code such as eval(), new Function(), os.system(), and child_process.exec(), as well as unsafe deserialization methods and browser injection patterns involving dangerouslySetInnerHTML and .innerHTML=.
The second stage runs after each model turn. At that point, Claude reviews the full git diff created during the session and looks for issues that pattern matching might miss. Anthropic said this stage can identify authorization bypass, insecure direct object references, injection flaws, server-side request forgery, and weak cryptography.
The third stage activates when Claude performs commits or pushes through its Bash tool. In that review, the system checks surrounding files, sanitizers, and related code paths to confirm findings and cut down on false positives.
Developers can also add custom rules and repository-specific security checks to all three layers. That makes the plugin more adaptable for teams with internal coding standards or extra controls around sensitive repositories.
Anthropic said it has already used the plugin internally. “Across our internal rollout and benchmarks, we’ve seen a 30-40% decrease in security-related comments on PRs opened using the plugin,” the company said. “The plugin serves as a lightweight first pass, catching issues before a full code review.”
The plugin is free for all users and available on all plans, according to Anthropic. Instant security checks do not use model calls and do not add usage costs, while deeper reviews use the same Claude usage budget as standard requests.
Anthropic said the plugin requires Claude Code version 2.1.144 or later and Python 3.8 or newer. The deeper review stages work only inside git repositories, while the lightweight pattern checks can run in any directory.
