unchore-defendSAFE
Security help for defenders: grade your site's security headers (fixes + README badge) and read logs/code to find what happened. Claude Code plugin, MCP server, no install.
Overview
From the repository's own README, as read at the audited commit. Badges and raw HTML are left out.
Security help for the person protecting the system. Two tools, no install, works in Claude Code, Claude Desktop, any MCP client, or a plain terminal.
English · 한국어
Part of [Unchore](https://github.com/smilemino/unchore-ai) — the open-source personal AI that offers to automate what you ask twice.
July 2026: OpenAI's test agents broke out of their sandbox and got into Hugging Face's production systems — about 17,600 attacker actions across 11 nodes over roughly 4.5 days. Last week Australia said OpenAI agents had also got into its Medicare statistics portal. When Hugging Face's responders asked commercial frontier models to help read the attack, the requests were "blocked by the providers' safety guardrails, which cannot distinguish an incident responder from an attacker." They finished the forensics on an open-weight model (GLM-5.2). Sources: disclosure · technical timeline · OpenAI's statement (NPR) · Australia (ABC) We re-ran the published attack artifacts. A major US model's content filter blocked 11 of 14 requests. Defend still got an answer to all 14 — results and a one-command rerun. Defend makes that switch for you: it asks Claude, then GPT, then GLM, then DeepSeek, and moves on only when one refuses.
5d633251d2b3OBSERVED · 2026-10-11Connect
Built from this server's own package name, version and transport as found in its source — not copied from anyone's documentation, so it cannot drift against a page we do not control. Replace the environment placeholders with a token scoped to the least it needs.
claude mcp add unchore-defend --env OPENROUTER_API_KEY=${OPENROUTER_API_KEY} --env UNCHORE_TOKEN=${UNCHORE_TOKEN} -- npx -y [email protected]{
"mcpServers": {
"unchore-defend": {
"command": "npx",
"args": [
"-y",
"[email protected]"
],
"env": {
"OPENROUTER_API_KEY": "${OPENROUTER_API_KEY}",
"UNCHORE_TOKEN": "${UNCHORE_TOKEN}"
}
}
}
}Exposed tools (2)
2 read · 0 write · 0 destructive.
| Tool | Risk | Description |
|---|---|---|
defend_analyze | read | Read a log, code, config or suspicious message and say what happened, how bad it is and what to fix first. Personal data and secrets are masked before sending. If one AI refuses, the next answers (Claude → GPT → GLM → DeepSeek). Needs UNCHORE_TOKEN or OPENROUTER_API_KEY. |
site_check | read | Grade a public website |
Trust audit
SAFEgrade B · trust 89/100 Nothing in the source contradicts what it says it does. Grade A is reserved for packages that have also passed the behavioural sandbox.
| Layer | What it checks | Result |
|---|---|---|
| L0 | Provenance & inventory | PASS |
| L1 | Static analysis of the code | PASS |
| L2 | Instruction surface (what it tells the agent) | PASS |
| L3 | Class-specific surface | PASS |
| L4 | Behavioural (sandbox) | SKIPPED |
What the source does
- Filesystem
- declared (1 observation(s))
- Network
- declared (3 observation(s))
- Shell
- none-observed
- Dependencies
- pinned
- Secrets in source
- none-found
Findings (6)
.mcpbignore
import { ask, CHAIN } from "../../skills/unchore-defend/scripts/defend.mjs";const r = await handle({ jsonrpc: "2.0", id: 5, method: "tools/call", params: { name: "site_check", arguments: { url: "169.254.169.254" } } });for (const a of ["localhost", "127.0.0.1", "10.0.0.5", "169.254.169.254", "[::1]", "[::ffff:127.0.0.1]", "intranet", "a.local", "ex.com:8080", "ftp://ex.com", "https://u:[email protected]", ""]) assert.equal(c
const hop = async (u) => { seen.push(u.href); return u.hostname === "ex.com" ? { status: 301, headers: { location: "http://127.0.0.1/" } } : { status: 200, headers: {} }; };docs/demo.gif
Gates applied: no_behavioural_pass.
5d633251d2b3full audit observations/trust-audit/mcp-server/smilemino__unchore-defend.json · Report an issue / request a re-scanAudit history
Every audit this server has had. A grade with a past is a grade somebody is still checking.
| Date | Source | Verdict | Grade | Score | Change |
|---|---|---|---|---|---|
| 2026-10-11 | 5d633251d2b3 | SAFE | B | 89 | first audit |
Questions
What is the unchore-defend MCP server?
Security help for defenders: grade your site's security headers (fixes + README badge) and read logs/code to find what happened. Claude Code plugin, MCP server, no install.
What tools does unchore-defend expose?
2 in total: 2 read-only, 0 that write, and 0 that can delete or overwrite. Every one is listed on this page with its risk.
Is unchore-defend safe to connect to an agent?
The audit found nothing in the source that contradicts what it says it does, and graded it B (89/100). Grade A is held back for packages that have also passed a sandboxed behavioural run, which is why a clean server reads B.
What credentials does unchore-defend need?
It reads OPENROUTER_API_KEY and UNCHORE_TOKEN from the environment. Give it a token scoped to the least it needs — an agent that can be talked into calling a tool can be talked into calling it with your credentials.
How does unchore-defend run?
It speaks stdio, so it runs as a local process your client starts. It is published on npm as unchore-defend at 0.1.0.
How current is this page?
The grade is for one exact copy of the source (5d633251d2b3), read on 2026-10-11. The repository is watched and re-audited when it changes.