ClawHub
Back to skill

Security audit

reel-watch

Security checks for vulnerabilities and agentic risk

Overview

This skill does what it claims: downloads or reads user-provided videos/images, analyzes them locally or with Gemini, and saves outputs under a reels folder.

Install only if you are comfortable with user-provided media, captions, and screenshots being sent to Google Gemini when a Gemini key is configured. Use --engine local or omit GEMINI_API_KEY for sensitive content, set REEL_HOME to a dedicated folder, and avoid REEL_IG_COOKIES unless you intentionally want authenticated Instagram fetching.

A.I.G

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs

SkillSpector

Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (11)

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The declared purpose centers on analyzing video content and extracting information from it. However, this code chunk contains only support logic for quota accounting via a local JSON file and lock file. It does not fetch videos, decode media, invoke Gemini or Whisper for analysis, or produce any report about video contents. While quota tracking could be a supporting utility in a larger video-analysis skill, this chunk by itself does not match the declared behavior and instead implements a materially different function.

Content

No source excerpt is available for this finding.

Credential Access

High
Category
Privilege Escalation
Confidence
89% confidence
Finding

The skill reads API keys and a path to Instagram cookies from the environment or a .env file and uses them in a workflow that fetches untrusted remote content. Even though it says other .env entries are ignored, exposing credential-bearing inputs to a broadly scoped skill increases the chance of accidental disclosure, misuse of authenticated sessions, or unauthorized requests under the user's identity.

Content

Scanner excerpt · SKILL.md (reported line 19)May include surrounding context.

md
{ "name": "REEL_IG_COOKIES", "required": false, "description": "Path to a cookies.txt, only used as a last resort for Instagram posts that need a login." },
            { "name": "REEL_GEMINI_MODEL", "required": false, "description": "Gemini model to try first." },
            { "name": "REEL_ENGINE", "required": false, "description": "auto (default), gemini or local." },
            { "name": "REEL_HOME", "required": false, "description": "Folder for reels/ and .env instead of the current folder." },
            { "name": "REEL_WHISPER_MODEL", "required": false, "description": "faster-whisper model for local transcripts (default small)." }
          ],
        "install":

Credential Access

High
Category
Privilege Escalation
Confidence
60% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · reel.py (reported line 18)May include surrounding context.

python
1. fetch   - free download chain: local file -> yt-dlp (no login) -> kkinstagram redirect
                 -> yt-dlp with cookies (only if REEL_IG_COOKIES points to a cookies.txt)
    2. gemini  - (main) Gemini watches the whole video with audio (or looks at the images) and returns
                 a breakdown + transcript. Needs GEMINI_API_KEY (env var, or GEMINI_API_KEY=... in a .env file in the current folder).
                 A few frames are still extracted so Claude can spot-check on-screen text.
    3. local   - (backup, used when Gemini is off or fails) ffmpeg frames + faster-whisper transcript
    4. output  - manifest.json + a readable summary on stdout

Credential Access

High
Category
Privilege Escalation
Confidence
60% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · reel.py (reported line 44)May include surrounding context.

python
1. fetch   - free download chain: local file -> yt-dlp (no login) -> kkinstagram redirect
                 -> yt-dlp with cookies (only if REEL_IG_COOKIES points to a cookies.txt)
    2. gemini  - (main) Gemini watches the whole video with audio (or looks at the images) and returns
                 a breakdown + transcript. Needs GEMINI_API_KEY (env var, or GEMINI_API_KEY=... in a .env file in the current folder).
                 A few frames are still extracted so Claude can spot-check on-screen text.
    3. local   - (backup, used when Gemini is off or fails) ffmpeg frames + faster-whisper transcript
    4. output  - manifest.json + a readable summary on stdout

Credential Access

High
Category
Privilege Escalation
Confidence
80% confidence
Finding

The code automatically reads DATA_ROOT/.env from the current working directory or REEL_HOME and imports API-related variables into the process without any trust boundary checks. In an agent or shared-workspace context, this can cause the tool to consume attacker-planted configuration or credentials from an untrusted directory, leading to unintended secret use, outbound requests under the wrong account, or processing with attacker-controlled settings.

Content

Scanner excerpt · reel.py (reported line 309)May include surrounding context.

python
def load_dotenv():
    """Read only GEMINI_*, GOOGLE_API_KEY and REEL_* settings from ./.env; everything else in it is ignored."""
    for env in (DATA_ROOT / ".env",):
        if not env.exists():
            continue
        for line in env.read_text(encoding="utf-8").splitlines():

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
93% confidence
Finding

The skill explicitly instructs the agent to use shell, network, filesystem, and environment-backed behavior, but it declares no tool/permission scope. That means an agent runtime may grant broader capabilities than users expect, increasing the chance of unintended downloads, file writes, and access to local secrets during execution. The risk is amplified because the workflow handles untrusted URLs and local paths.

Content

No source excerpt is available for this finding.

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · reel.py (reported line 118)May include surrounding context.

python
"--write-info-json", "-o", str(work / "video.%(ext)s"), url]
    if cookies:
        cmd[1:1] = ["--cookies", cookies]
    r = subprocess.run(cmd, capture_output=True, text=True, timeout=600)
    vids = [p for p in work.glob("video.*") if p.suffix not in (".json", ".part")]
    if r.returncode != 0 or not vids:
        raise RuntimeError(r.stderr.strip().splitlines()[-1] if r.stderr.strip() else "yt-dlp failed")

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · reel.py (reported line 214)May include surrounding context.

python
# ---------------------------------------------------------------- media
def duration_of(ff, video):
    r = subprocess.run([ff, "-i", str(video)], capture_output=True, text=True, encoding="utf-8", errors="replace")
    m = re.search(r"Duration: (\d+):(\d+):([\d.]+)", r.stderr)
    return int(m.group(1)) * 3600 + int(m.group(2)) * 60 + float(m.group(3)) if m else None

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · reel.py (reported line 224)May include surrounding context.

python
fdir.mkdir(exist_ok=True)
    n = max(6, min(max_frames, math.ceil(dur / 2))) if dur else max_frames
    fps = n / dur if dur else 0.5
    subprocess.run([ff, "-loglevel", "error", "-i", str(video), "-vf", f"fps={fps:.5f},scale=720:-2",
                    "-q:v", "3", str(fdir / "%03d.jpg")], check=True)
    frames = sorted(fdir.glob("*.jpg"))
    return [{"t": round((i + 0.5) / fps, 1), "path": str(f)} for i, f in enumerate(frames)]

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · reel.py (reported line 232)May include surrounding context.

python
def transcribe(ff, video, work, model_name):
    wav = work / "audio.wav"
    r = subprocess.run([ff, "-loglevel", "error", "-i", str(video), "-vn", "-ac", "1", "-ar", "16000", str(wav)])
    if r.returncode != 0 or not wav.exists():
        return None
    try:

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

When Gemini is enabled, the skill uploads user-provided videos, images, and potentially captions/metadata to Google's Gemini service automatically, but the execution path does not present a clear runtime warning or require explicit user consent at the point of transfer. In a skill designed to inspect arbitrary user-shared media, this creates a real privacy and data-handling risk because sensitive content may be sent to a third party unexpectedly.

Content

No source excerpt is available for this finding.

Static analysis

No suspicious patterns detected.