Backed by real research — a published CVE on a major MCP server

Secure the servers
your agents trust.

Gridsline scans MCP servers — the code that connects AI agents to tools, files, and credentials — and uses an LLM to verify every finding. You get confirmed vulnerabilities, not noise.

Request access How it works
gridsline — scan Example output

$ gridsline scan ./mcp-server

Parsed 142 files · 38 tool handlers

▲ 23 candidate findings

Verifying with LLM…

✕ 19 dismissed — not reachable or not exploitable

● 4 confirmed

HighPath traversal in read_filesrc/tools/fs.ts:88 HighCommand injection in run_tasksrc/tools/exec.ts:41 MedSSRF in fetch_urlsrc/tools/http.ts:23 LowToken leaked in error outputsrc/server.ts:112

The problem

Agents got the keys.
Nobody checked the locks.

Agents are wired in fast

Teams are connecting AI agents to file systems, databases, internal APIs, and production credentials — often through MCP servers pulled straight from GitHub.

MCP servers are wide open

Every server sits between untrusted model input and real side effects. Path traversal, command injection, SSRF, and leaked secrets are one tool call away.

Scanners cry wolf

Generic static analysis buries real issues under false positives. When everything is flagged, nothing gets fixed — and the one that matters ships.

How it works

Scan. Verify. Disclose.

Three steps from source code to a fixed, published vulnerability.

  1. 01

    Scan

    We read the server's source — every tool handler, every input path, every place untrusted data reaches the file system, a shell, or the network — and flag candidate vulnerabilities.

  2. 02

    AI-verify

    An LLM reviews each candidate in context, tracing whether it's actually reachable and exploitable. Anything that can't be triggered is dropped. What's left is real.

  3. 03

    Disclose

    Confirmed issues go privately to the maintainers first. Once fixed, they're added to our public dataset of verified MCP vulnerabilities.

Why Gridsline

Signal, not noise.

A public dataset of verified MCP vulnerabilities

Every confirmed finding — after the fix ships — goes into an open record the whole ecosystem can learn from. No guesses, no unconfirmed scanner output.

Verified before you see it

LLM verification removes the false positives, so every issue that reaches a maintainer is worth their time.

Responsible disclosure

Maintainers hear from us first, privately, with a clear write-up and a coordinated timeline to fix it.