Semgrep Agentic AppSec Platform: scan agent code pre-commit
The Semgrep Agentic AppSec Platform puts a security scanner inside Claude Code, Cursor and Codex through an MCP server. The numbers are vendor benchmarks. Test it on your repo.
The Semgrep Agentic AppSec Platform moves the security scan from the pull request to the moment the agent writes the code. Semgrep launched it on November 12. Coding agents in Claude Code, Cursor and Codex can call the scanner through an MCP server and fix a finding before the commit. That is the correct place for the check. The performance numbers are a different matter: Semgrep measured all of them, and it published no method.
What actually happened
Semgrep gave the details in its launch post and a press release. The platform has five parts:
- Guardian. Checks code while the agent writes it, in any coding harness, before the commit. Semgrep says it now uses AI detection for authorization, IDOR and business logic flaws.
- Agentic Workflows. Security agents that hunt for authentication, authorization, injection and business logic bugs that pattern matching does not find. They now also run in CI pipelines.
- Mandoline. A program analysis tool. It traces control flow and data flow from the entry points and returns only the code paths that an attacker can reach.
- Semgrep MCP server. Gives agents in Codex, Claude Code, Cursor and other MCP tools access to the platform.
- Context layer. A stored knowledge graph of your code facts, documents and developer feedback. This part is in private beta.
Guardian, Agentic Workflows, Mandoline and the MCP server are available now to Semgrep customers. Neither source states a price. You can supply your own model key, and there is a single-tenant option and an EU region.
The claims, all from Semgrep's own benchmarks:
- Agents with the platform found more than twice as many true vulnerabilities as frontier models alone.
- Mandoline cut token cost in security workflows by 65% on average.
- Mandoline scored 100% on the OWASP Benchmark for Java and 98.28% for Python, in one run with no AI tokens.
Semgrep did not publish the dataset, the baseline or the method for the first two. The OWASP score uses a public scorer, so that one you can check.
Why the Semgrep Agentic AppSec Platform matters for your business
Your developers now accept more code than they can read. Review is the slow step, and agents made the queue longer. A scan at pull request time finds the bug after the agent has built three more files on top of it.
The design idea here does not depend on Semgrep. Give the agent a deterministic tool, and let the model reason on the result. A model is good at intent. It is not reliable when it must trace one input through 200 files. A program analysis tool does that in milliseconds and costs no tokens. We made the same point when Black Duck shipped agent-written fixes: the fix still needs a person to approve it.
What to do:
- Put one scanner in the agent loop. Semgrep's open-source CLI, CodeQL or a different tool. The agent must run it before each commit.
- Keep the CI scan. The pre-commit check is fast. The CI check is the gate. Use the two together.
- Test the claims on your own repo. Take ten bugs you already fixed. Count how many the tool finds, with and without the agent.
- Scope the MCP server. It reads your full codebase. Give it a token with the minimum access, and know where the findings go.
- Measure tokens per task. If the 65% number is true for you, your bill will show it in a week.
Key takeaways
- Semgrep launched its Agentic AppSec Platform on November 12: Guardian, Agentic Workflows, Mandoline, an MCP server and a context layer in private beta
- Coding agents in Claude Code, Cursor and Codex can call the scanner through the MCP server before the commit
- The "more than twice" and "65% less token cost" numbers are Semgrep benchmarks with no published method
- No price is stated; the platform is available to Semgrep customers
- Put a deterministic scanner in the agent loop, keep the CI gate, and test the claims on bugs you already know
Your agents write code faster than your team can review it. We set up coding agents with a scanner in the loop, a CI gate and a human sign-off step, so the speed does not become security debt. Ask us to review your agent workflow, or see how we build software you own.
Sources: Semgrep, Business Wire via Yahoo Finance.
- #semgrep
- #appsec
- #mcp
- #coding-agents
- #code-security
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Marathon Health data breach: a fake IT call got the password
In the Marathon Health data breach, a caller posed as IT support and got one employee's password. Data left a cloud app in hours. Verify each IT support call.
Read itTSMC employee data breach traced to a cloud provider
TSMC confirmed an employee data breach in an SEC filing. The leaked data matches records at a cloud service provider. Inventory the HR data your vendors hold.
Read it