arkx · advanced · level 7
Arkx — Master of Resources
| Target | Arkx · Advanced track · Level 7 |
|---|---|
| Vulnerability | MCP resources/tools exposed behind an Origin-only (DNS-rebinding) guard |
| OWASP LLM Top 10 | LLM06 Excessive Agency · LLM02 Sensitive Information Disclosure |
| MITRE ATLAS | AML.T0057 LLM Data Leakage · AML.T0053 LLM Plugin Compromise |
| Flag | flag-5a7c89 |

Result
The level claims its “main interface is down”. It isn’t the whole story: the agent is backed by a Model Context Protocol server (Python MCP SDK, Streamable HTTP, serverInfo.name = "Weather") reachable at /level/7/mcp. Its only protection is a DNS-rebinding Origin check — which exists to stop browsers, not clients.
Flag: flag-5a7c89
Exploit
1 — The insight. A browser fetch() always sends an Origin header → the server rejects it (403, DNS-rebinding guard). Requests with no Origin are allowed. A browser can’t drop its own Origin, so you need a non-browser HTTP client. (The level’s normal session cookies authenticate the request; used only against the target, for this one authorized call.)
2 — Handshake. POST /level/7/mcp with Accept: application/json, text/event-stream, the JSON-RPC initialize body, and no Origin header → 200, and the server hands back an Mcp-Session-Id. Then notifications/initialized (202) and resources/list:
resource://flag → "the flag is locked"
resource://super-secret → (a rick-roll decoy)
3 — Enumerate tools. tools/list → a single tool, unlock_flag (no arguments).
4 — Exploit. Call it, then read the resource:
tools/call unlock_flag → "Flag has been unlocked"
resources/read resource://flag → "You did it! the flag is: flag-5a7c89"
Root cause
A sensitive resource and the tool that unlocks it are published on an MCP server. MCP is designed to expose its resources and tools to any client that can reach it — and the only barrier here, an Origin/DNS-rebinding check, is an anti-CSRF measure, not an authorization control. There is no per-resource auth at the MCP layer.
Impact
Anyone who can reach the endpoint reads “protected” resources and invokes privileged tools, bypassing the UI and its auth entirely. Takeaway: protected data should never live on an MCP surface, and MCP tools/resources need real authorization — the transport guard is not it.