Semgrep Guardian, our security plugin for AI coding agents, now runs AI-based detection, deploys org-wide policies to every developer from one place, and continuously prevents your agent from making risky supply chain changes.
Teams adopt coding agents to keep up with the pace of software, and those agents now produce more code than any security team can review by hand. Attackers run agents of their own, around the clock, against the same applications. Scanning after the merge arrives too late for that volume, and the agents writing the code often lack security tools built for them. Guardian puts those tools next to the agent. It checks code as your agent generates it so your team has less to triage in PRs & production.
Guardian is the part of Semgrep's agentic code security suite that sits next to the agent while it writes.
Guardian Policies are generally available
Guardian Policies are now generally available (GA). You set a policy once, and it runs everywhere your developers use Guardian.
A policy lets you do two things:
Add your own custom rules, so Guardian checks for the patterns specific to your codebase and your standards.
Change which rules run for your developers, so the checks match your risk.
For example, you can add a custom rule that flags calls to a deprecated internal authentication helper. Every agent in your organization then checks for it.
Developers can write custom rules too. Semgrep rules look like the code they match, so a rule for your framework or an internal API is straightforward to write. Test it locally, add it to your local rules file, and it applies to your future scans. Or put it in a repo-wide rules file so your entire team benefits.
The point -- Security teams get one place to define what secure means. Developers get the same checks no matter which agent or editor they use.
Guardian AI Detection finds what rules can't (public beta)
Static rules are good at patterns you already know, but some problems don't have a pattern you can write in advance. Guardian AI Detection, now in public beta, runs a second, adversarial agent alongside your coding agent and reviews the code it produces the way an attacker would.
It looks for authorization flaws such as insecure direct object references (IDOR) and broken object-level authorization (BOLA), which let one user read or change another user's data. These are hard for static rules to catch because the code looks correct & (often) the problem is a check that isn't there.
Guardian also looks for malicious intent. Supply chain attacks like Shai-Hulud showed us what that looks like: code that quietly steals credentials, or a CI change that uploads data to a remote server. Either way, Guardian has you covered.
Your agent now has static and AI-based scanning, all in the same place.
Supply chain advisories and lockfile detection
Supply chain incidents arrive faster than most teams can read about them, let alone respond to them. Guardian now takes on part of that work.
If your agent tries to change a dependency in your lockfile in a way that introduces malware or a reachable vulnerability (one your code can actually call), Guardian blocks the change and flags it, so it never reaches your build system or production. A lockfile pins the exact dependency versions your project installs, so an unexpected change there changes what you ship.
For major security incidents, Guardian also pushes out an advisory and checks that the versions installed locally get upgraded. That covers incidents like PHP2Shell, a PHP runtime remote code execution (RCE) vulnerability, and Node2Shell. Your developers get the advisory, and Guardian makes sure their local versions are properly upgraded.
Ask your agent about Semgrep App
Guardian now comes with an MCP (Model Context Protocol) server, in public beta, that connects your agent to Semgrep App. It's pre-installed when you install Guardian, so there's nothing extra to set up. Your agent can then work with your Semgrep data directly:
Detect: update your policies and push new rules to your Semgrep account.
Fix: pull your existing findings and kick off a campaign to fix them.
Report: summarize what's been fixed and the major risks in your codebase, then report upwards.
All without leaving your agent.
Deploy Guardian across your enterprise
Rolling Guardian out across a company raises one question fast: how well is it actually deployed? New reporting gives you the visibility to know, so you can see how Guardian is running across your codebase and where it isn't yet.
In the early stages of a pilot, deploy Guardian in observe-only mode. It silently collects metrics in the background, so you can build that picture before you ever surface a finding to a developer.
Guardian also now supports a new harness, Copilot, alongside Codex and Claude Code, so your developers can stay in the agent they already use.
Get started with Guardian
In Claude Code, run:
claude plugin install semgrep@claude-plugins-officialThen start a new session with claude and sign in to Semgrep when prompted.
In Codex, add the Semgrep marketplace and install the plugin:
codex plugin marketplace add https://github.com/semgrep/guardian --sparse .claude-plugin --sparse plugin
codex plugin add semgrep@semgrep-marketplaceThen start a new session with codex and ask Codex to log in to Semgrep Guardian. The full steps are in the Guardian quickstart.
In Copilot (Public Beta)
add the Semgrep plugin marketplace to Copilot:
copilot plugin marketplace add https://github.com/semgrep/guardianinstall the Guardian Copilot plugin from that marketplace:
copilot plugin install guardian-copilot@semgrep-marketplace
When you next open the Copilot CLI, Guardian will prompt the agent to log in via MCP tools after your first prompt of the session -- or you can ask the agent “log into Semgrep Guardian via the MCP”.
Rolling it out company-wide? Deploy Guardian across your organization through your mobile device management (MDM) tool, and reach out to us for support.
