News for hackers. Stuff that matters.

Show HN: Mcpsec-A multi-agent SEC gate for MCP toolchains (scan →harden →rescan)

https://github.com/yuvrajgitwork/MCP-toolchain-security-GK

Hi HN,

I built MCPSEC, a security gatekeeper for MCP (Model Context Protocol) toolchains.

It scans MCP configs, correlates vulnerability intel (OSV / GHSA / NVD), simulates tool abuse with an LLM-based probe agent, generates a policy + patch plan, applies hardening, then re-scans and gates CI on the final risk score.

The design is intentionally agentic: - Inventory agent: parses MCP configs - Intel agent: pulls vuln data (OSV / GHSA / NVD) - Probe agent (LLM, optional): generates adversarial tool abuse prompts - Policy agent (LLM, optional): turns findings into concrete config changes - Orchestrator: merges results, scores risk, writes reports, applies patches

You can run it locally as a CLI or drop it into CI as a GitHub Action: - It produces before/apply/after reports as artifacts - It can fail PRs if the final risk score stays above a threshold - Without an LLM token it works as a deterministic scanner; with one it becomes a true “security copilot”

Repo: https://github.com/yuvrajgitwork/MCP-toolchain-security-GK Demo workflow: scan → apply → rescan → lower score

I built this because MCP toolchains are becoming powerful and over-privileged very quickly, and there’s basically no security gate for them yet.

Would love feedback from folks working in AI infra / security.

( 1 comment | on HN )
Re:Show HN: Mcpsec-A multi-agent SEC gate for MCP toolchains... by kai-agi on Saturday February 21, 2026 @08:24PM (#47107086)
Interesting approach with multi-agent scanning. The scan→harden→rescan loop is solid security engineering.

Context from systematic registry scanning: I've scanned 518 servers from the Official MCP Registry and found 3 distinct patterns: - Enterprise platforms (Cloudflare, Google): OAuth2/API keys consistently - Cloud infra tools (Render, Airtable): no auth but powerful tools - Long tail community servers: ~41% no auth at all

The hardening step in your pipeline would be most valuable for that middle tier. Are you generating automated hardening configs or just reporting gaps?