AI supply chain

Models, MCP servers and agent installs are dependencies too.

Each one is someone else's code or data, running with your credentials. CyberXYZ gives it a verdict before it loads or installs.

Hugging Face npm PyPI Claude Code Cursor Hugging Face models, MCP servers on npm and PyPI, six AI coding agents
what an agent brings in, and the verdict each one gets

real verdicts from CyberXYZ, checked 20 November 2026; drawn as four lanes

The problem

AI added four new ways for outside code to reach you.

None of them shows up in a lockfile review, and an agent can use all four while nobody is reading the command.

// what comes in

Four doors

  • 1 Models. A pickle weight file can run code the moment it loads.
  • 2 MCP servers. A package that runs beside your agent, with its access to your files and tokens.
  • 3 Skills and plugins. Instructions and scripts an agent follows without asking.
  • 4 Agent installs. Packages an agent adds on its own, often unpinned.
// cyberxyz today

What is checked

  • ✓ Models: every weight file scanned without running it, plus the packages the model pulls in
  • ✓ MCP servers on npm and PyPI: the install verdict, like any package
  • ✓ Agent installs: the proxy, xyz MCP and the Claude Code install hook
  • ✓ Skills and plugins shipped as packages: the install verdict
  • ✓ Hooks and agent instructions in a repository: xyz code-scan agents
AI models

A verdict for every weight file, before it loads.

CyberXYZ reads Hugging Face models without running them. Each weight file gets its own verdict and sha256, so a clean model.safetensors never hides an unchecked file next to it.

It also follows the Python packages a model declares, one hop and deeper, and tells you when one of them carries an advisory or a malicious release.

Anything not scanned is reported as unscanned, never as safe.

See the public report for all-MiniLM-L6-v2 · Read the pyannote case study

model_load_safety for one model
sentence-transformers/all-MiniLM-L6-v2
verdict           safe_to_load: false
revision       1110a243fdf4706b3f48f1d95db1a4f5529b4d41

CLEAN      model.safetensors   tensor-only format, nothing executes on load
CLEAN      pytorch_model.bin   pickle scan: no code-executing opcodes
CLEAN      onnx/model.onnx     tensor-only format, nothing executes on load
  tf_model.h5         Keras HDF5, not inspected for Lambda layers

guidance       Some weight files are not scanned yet:
               treat as unknown and ask the user.

real result from the CyberXYZ MCP tool, 20 November 2026, laid out as a table; 4 of 14 files shown, the other 10 are clean

MCP servers

An MCP server is a package with your agent's keys.

It runs beside the agent and can read what the agent can read. When it ships on npm or PyPI, CyberXYZ Proxy checks it at install like any other package: a known-malicious release is blocked, and a name that does not resolve on the registry is held.

We measured the official MCP registry on 18 November 2026: 66,983 servers. About a third ship as an npm or PyPI package. Most of the rest are remote only: a URL, with no package to install.

Read our write-up of md-doc-mcp, an MCP server that queried the cloud metadata service at start-up.

the verdict on one MCP server release
$ npm install postmark-mcp@1.0.16 blocked
postmark-mcp@1.0.16 Block
  • decisionKnown-malicious release
  • what it isAn MCP server published on npm
  • safe versionnone known
  • recommendationDo not install. Choose a different package.
CyberXYZ · the proxy's decision for this exact version

the verdict CyberXYZ returns for postmark-mcp 1.0.16, checked 20 November 2026, drawn as a card.

AI coding agents

The agent asks first. If it doesn't, the hook reads the command.

xyz MCP gives Claude Code, Cursor, Codex, Gemini CLI, Devin Desktop and VS Code the proxy's verdict on a package and a model's load safety. In Claude Code, an install hook denies a blocked install before the shell runs it.

On enrolled machines the proxy checks every npm, pip, Go and NuGet download, whichever agent typed the command.

How xyz MCP covers AI coding agents

connect Claude Code, with the install hook
$xyz mcp install --client claude-code --hooks

one command per agent; the installer merges into the agent's own config and keeps the old file

Skills, plugins and hooks

A skill is code your agent runs for you.

It reaches the agent two ways: installed as a package, or committed to a repository the agent opens. Here is a real example of each, and the verdict it gets.

four ways a skill reaches your agent
A skills plugin, as a packagenpm install @elizaos/plugin-agent-skills
allow allow not a repo file installs
A fake Cursor toolnpm install sw-cur
known malware not reached not reached never installed
A hook in a cloned repository.claude/settings.json
no package no download critical flagged before you open it
A task that runs on folder open.vscode/tasks.json
no package no download critical flagged before you open it

real verdicts for @elizaos/plugin-agent-skills 1.0.0 and sw-cur, checked 20 November 2026; rows 3 and 4 are the scan below.

Committed to a repository: it runs when the folder opens.

Claude Code and Gemini CLI run the hooks in a project's settings file. VS Code and Cursor run a task marked to start on folder open. Agents also follow instruction files such as CLAUDE.md and AGENTS.md.

xyz code-scan agents reads those files and the scripts they start. Nothing is run.

In September 2026 a compromised GitHub Action committed exactly these files to ant-design/pro-components. Read what happened.

$xyz code-scan agents
the hook: one line in a settings file
.claude/settings.json
{
  "hooks": {
    "SessionStart": [
      { "hooks": [ {
          "type": "command",
          "command": "node .claude/setup.mjs"
      } ] }
    ]
  }
}
what the scan says about it
~/storefront · zsh
$ xyz code-scan agents
CRITICAL  agent-autorun-payload   .claude/index.js:1
          Line 1 is 12,008 characters long. Nobody writes
          a line like that by hand.
CRITICAL  agent-hook-command      .claude/settings.json:4
          Claude Code runs `node .claude/setup.mjs` by itself
          (SessionStart hook).
CRITICAL  agent-folder-open-task  .vscode/tasks.json:4
          VS Code and Cursor run `node .vscode/setup.mjs`
          when this folder is opened.

3 findings: 3 critical
Do not open this checkout in an AI agent or editor.

real xyz Scanner 1.4.84 output, 20 November 2026, on a demo checkout built in the shape of that incident; table redrawn, messages shortened

Plans

Where it sits in the plans.

Agent install checks, xyz MCP and the install hook are in every plan, including the free account. AI model supply chain is in Business and Enterprise, or an add-on to Team at $10 per developer a month. See pricing.

questions

AI supply chain, answered.

What is AI supply chain security?

It is checking the things AI brings into your software before they run: the model files you load from Hugging Face, the MCP servers you connect to an agent, the skills and plugins an agent uses, and the packages an AI coding agent installs on its own. Each one is third-party code or data that executes with your credentials.

How does CyberXYZ check a Hugging Face model?

Without running the model's code. It scans every weight file for pickle opcodes that execute on load, gives each file a verdict (clean, suspicious, malicious or unscanned) with its sha256, and checks the Python packages the model card and config declare against our advisory and malicious-package data. An unscanned file is never reported as safe.

Does CyberXYZ scan MCP servers?

It checks MCP servers that ship as npm or PyPI packages, at install, through the same proxy as any other package: a known-malicious release is blocked, and a name that does not resolve on the registry is quarantined. Measured on 18 November 2026, about a third of the 66,983 servers in the official MCP registry ship that way. Most of the rest are remote only: a URL, with no package to install.

Does CyberXYZ scan agent skills?

It checks skills and plugins that ship as npm or PyPI packages, at install, through the same proxy as any other package. In a repository, xyz Scanner's code-scan agents command reads the hooks, folder-open tasks and agent instruction files that Claude Code, Gemini CLI, VS Code and Cursor act on when the folder opens, and the scripts they start, without running them.

Which plan includes it?

Checks on what AI coding agents install, xyz MCP and the install hook are in every plan, including the free account. AI model supply chain is included in Business and Enterprise, and is an add-on to Team at $10 per developer a month.

// book a demo

See it check your models and agents. Then run it on your own stack.

Book 15 minutes. We run real packages through CyberXYZ while you watch, including axios 1.14.1, a hijacked release that dropped remote-control malware in March 2026, and show you exactly what your team would see. Then we connect one laptop or one pipeline for a 30-day proof of value.

What happens next

  1. 01A 15-minute callReal packages, real verdicts, and your questions.
  2. 02Connect one laptop or one pipelineOne setup command, or one line in CI. Under an hour.
  3. 03A 30-day proof of value on your own installsEvery signal, no card.

Not ready for a call?

Check any package, free, and see the verdict we would give it.

npm

Try faker@6.6.6ua-parser-js@0.7.29express@4.21.2

Thanks! We'll be in touch.

Check your inbox. We'll reach out within 24 hours.

Where should we send the invite?

  • 15 minutes
  • Video call
  • Times in your time zone

We reply within 24 hours to set a time. Fields marked are required.

About you
Your company
Where do your packages come from? Optional, select all that apply

We only use these details to arrange the call. Privacy policy