Back to skill

Security audit

clipping-agency-skill

Security checks for vulnerabilities and agentic risk

Overview

ClawScan marked this skill suspicious because scanners report risks involving RSS input, upload destinations, API credentials, and approval handling; Skillspector coverage was partial, so the evidence is incomplete.

Consider the suspicious verdict carefully before installing. In particular, review whether the RSS, upload, credential, and silence-based approval behavior is acceptable for your use. Skillspector inspected all listed components, but marked its analysis partial because some local references were missing or ambiguous; the overall evidence is therefore incomplete.

Vulnerability Patterns
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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
Findings (1)

T09 · Insecure Skill Coding Practices

Warning
Location
scripts/sources.py
Finding

Untrusted RSS enclosures can access local files and internal network services

Content
View full analysis

Impact

A malicious enclosure can copy any known, process-readable regular file into the episode directory as source.*, or cause an HTTP GET to a reachable loopback, private-network, or link-local service and save its response. This can introduce unrelated sensitive data into the client workspace and reach services unavailable to the feed publisher. Remote retrieval of the saved data, automatic exfiltration, and code execution were not established.

Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
Findings (47)

Harmful Content Injection

Critical
Category
Prompt Injection
Confidence
90% confidence
Finding

This content may contain harmful instructions that could cause physical harm if followed. CRITICAL: Review carefully before use.

Content

Scanner excerpt · scripts/check_clips.py (reported line 343)May include surrounding context.

python
path = clip["file"] if os.path.isabs(clip["file"]) else os.path.join(bdir, clip["file"])
            ranges = clean_ranges(clip.get("ranges")) or (
                [(float(clip["start"]), float(clip["end"]))] if is_num(clip.get("start"), 0) and is_num(clip.get("end"), 0)
                and clip["end"] > clip["start"] else None)
            if ranges is None:
                add("error", "a clip you cut yourself needs \"start\" and \"end\" (or \"ranges\") in seconds of the episode")
                ranges = []
        elif m:
            path = os.path.join(bdir, "clips", m.get("file") or f"clip-{k}.mp4")
            ranges = clean_ranges(m.get("ranges")) or []
        else:

Tainted flow: 'req' from os.environ.get (line 178, credential/environment) → urllib.request.urlopen (network output)

Critical
Category
Data Flow
Confidence
90% confidence
Finding

Credentials or environment variables flow to a network sink. This is a high-confidence indicator of credential exfiltration.

Content

Scanner excerpt · scripts/transcribe.py (reported line 181)May include surrounding context.

python
req = urllib.request.Request(url, data=bytes(body), method="POST", headers={
        "Authorization": f"Bearer {key}", "Content-Type": f"multipart/form-data; boundary={boundary}"})
    try:
        with urllib.request.urlopen(req, timeout=600) as r:
            return json.loads(r.read())
    except urllib.error.HTTPError as e:
        die(f"transcription API error {e.code}: {e.read().decode(errors='replace')[:400]}")

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
88% confidence
Finding

Most visible behavior supports the declared approval workflow. However, page_rule(), silence_due(), and decisions() implement a substantive additional approval mode: unchanged clips can receive an automatic approve decision after the deadline without any client answer. This mode requires a configured written-agreement reference, is disclosed on the generated page, and stops when a reply is recorded, but the declared description does not mention it and instead describes client answers and approved-only posting. The supplied script does not itself publish posts or execute remote cancellations; it records approval eligibility and prepares cancellation requests. Other parts of the broader declared workflow cannot be assessed from this chunk alone.

Content

No source excerpt is available for this finding.

Context Leakage

High
Category
Data Exfiltration
Confidence
85% confidence
Finding

Code or instructions that leak agent conversation context to external services, potentially exposing sensitive user interactions.

Content

Scanner excerpt · SKILL.md (reported line 316)May include surrounding context.

md
- `scripts/schedule.py`: the schedule, `--media`, `--check`, `--posted`, `--cancelled`, `--status`.
- `scripts/status.py`: what waits on whom, every client.
- `scripts/report.py`: the monthly report.
- `scripts/upload_parts.py`: uploads a clip to a WoopSocial upload session and records its SHA-256.
- `scripts/agencylib.py`: shared helpers, platform limits, the ledger, schedules.
- `references/`: client-file.md, batch-file.md, clip-rubric.md, approval.md, platforms.md,
  woopsocial.md, monthly-report.md.

YARA rule 'agent_skill_mcp_tool_poisoning_metadata': MCP/tool metadata poisoning indicators in tool schemas or skill manifests [agent_skills]

High
Category
YARA Match
Confidence
80% confidence
Finding

YARA rule matched a hack tool or exploit indicator (offensive tools, reconnaissance, privilege escalation, or exploit frameworks).

Content

Scanner excerpt · scripts/approve.py (reported line 526)May include surrounding context.

python
ext.splitlines() if not l.lstrip().startswith(">"))


def normalize_reply(text):
    """The reply as plain lines: the same characters whatever app it was typed in (non-breaking and full-width
    spaces, full-width colons, curly quotes, en and em dashes, bullets, markdown, zero-width marks)."""
    t = unicodedata.normalize("NFKC", text.replace("\r\n", "\n").replace("\r", "\n"))
    t = re.sub(r"[​-‍⁠]", "", t)
    t = re.sub(r"[‘’‚‛′`´]", "'", t)
    t = re.sub(r"[“”„‟«»″\"]", "", t)
    t = re.sub(r"[‐-―−﹘﹣]", "-", t)
    t = re.sub(r"[ \t  -    ]+", " ", t)
    t = re.sub(r"\*\*|__|~~|\*", "", t)
    t = re.sub(r"(?m)^\s*(?:[-+>•·◦▪▫‣⁃●○■□]\s+)+", "", t)
    mt = CUT_RE.search(t)   # the message this reply answers, or a signature, starts here
    return t[:mt.start()] if mt else t


def _scan(text, n, allow):
    """Words in a stretch of reply text that make an answer unclear. allow: the classes that may appea

Some refused replies leave approval-on-silence active, contrary to the documented guarantee.

High
Category
Excessive Agency
Confidence
99% confidence
Finding

Several reply-rejection paths call die() directly instead of refuse(), including when the responder's identity is missing or does not match an approver. These paths do not persist a hold, so an otherwise eligible approval-on-silence page can still automatically approve clips despite a received objection. In a workflow that publishes to client social accounts, this can result in unwanted public posts and contradicts the documented fail-closed guarantee.

Content

No source excerpt is available for this finding.

Anti-Refusal Statement

High
Category
Anti-Refusal
Confidence
80% confidence
Finding

Skill instructs the agent to omit warnings, disclaimers, or ethical commentary. Stripping safety caveats hides risk from the user and is a common jailbreak preamble.

Content

Scanner excerpt · scripts/check_clips.py (reported line 23)May include surrounding context.

python
own words for that stretch. It warns on any word that isn't heard (a word cut out inside the clip),
on a first word that isn't the clip's first planned word (cut off), and on a last word after the last
planned one (the next sentence starting). A transcriber mishears old or noisy audio, so a warning means
listen to that spot; no warning is a strong hint, never a guarantee. The result is kept in
clips/listen.json per file, so it runs once per cut. AGENCY_LISTEN_MODEL picks the model (default
base; small hears better). Without faster-whisper it says so and skips this check; listen to each
clip yourself then.

Credential Access

High
Category
Privilege Escalation
Confidence
70% confidence
Finding

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

Content

Scanner excerpt · scripts/client.py (reported line 583)May include surrounding context.

python
accs = accounts_of(c)
    if still:
        lines.append("On the agency's side, for each account: `social_accounts_delete` with its id (this removes "
                     "WoopSocial's access token), then `client.py offboard "
                     f"{cid} --revoked <PLATFORM>`:")
        lines.append("")
        for p in still:

Silence can be converted into approval and used to schedule clips without an affirmative client response.

High
Category
Excessive Agency
Confidence
96% confidence
Finding

The manifest promises posting only approved clips, and L014-L015 describe approval for exactly the files and captions the client accepted. L148-L155 explicitly record what silence approved; those clips remain in the go list and become posting requests at L189-L204, allowing the affirmative approval boundary to be replaced by an automatic decision.

Content

No source excerpt is available for this finding.

Context Leakage

High
Category
Data Exfiltration
Confidence
85% confidence
Finding

Uploading the selected clip is intended, and the matched docstring alone is harmless, but load_session accepts any HTTPS upload destination from the session JSON without verifying its host or provenance. An attacker able to replace or influence that JSON can redirect confidential client-video bytes to an attacker-controlled server; multipart-count validation and the recorded hash do not authenticate the recipient.

Content

Scanner excerpt · scripts/upload_parts.py (reported line 2)May include surrounding context.

python
#!/usr/bin/env python3
"""Upload one file to a WoopSocial media upload session (standard library only).

WoopSocial uploads work in three steps:
  1. The agent calls `media_uploads_create_session` with the projectId and the file size in bytes.

Skill declares no tool scope ('permissions' or 'allowed-tools') but code capabilities were detected: env, file_read, file_write, mcp, network, shell.

Medium
Category
MCP Least Privilege
Confidence
70% confidence
Finding

Without declared permissions the skill's intent is opaque and cannot be validated.

Content

No source excerpt is available for this finding.

Context Window Stuffing

Medium
Category
Memory Poisoning
Confidence
80% confidence
Finding

Skill attempts to fill the context window with filler content, displacing legitimate instructions and safety constraints. This can degrade agent performance or bypass safety boundaries.

Content

Scanner excerpt · examples/agency/clients/example-orbit-podcast/episodes/ep-012/approval/approve-ep-012.html (reported line 70)May include surrounding context.

html
p4"><img src="data:image/jpeg;base64,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

The upload and posting workflow lacks a warning about external disclosure and public publication of client media.

Medium
Category
Output Handling
Confidence
86% confidence
Finding

L16 instructs the agent to upload client video files, and L24 instructs it to create scheduled social posts. Although the document labels the accounts as examples and lists the destinations, it does not warn that executing this workflow transfers media to external services and can make it publicly accessible; the example disclaimer is not a dry-run safeguard.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · references/woopsocial.md (reported line 32)May include surrounding context.

md
|---|---|
| Claude (claude.ai, Desktop, Cowork) | Add WoopSocial from the [Claude directory](https://claude.ai/directory/woopsocial), click Connect and sign in |
| Cursor | Install from [cursor.directory](https://cursor.directory/plugins/woopsocial-social-media-scheduler) and sign in |
| Claude Code | `claude mcp add --transport http woopsocial "https://api.woopsocial.com/mcp"`, then restart with `claude --continue`, `/mcp`, select woopsocial, Authenticate |
| Codex | `codex mcp add woopsocial --url https://api.woopsocial.com/mcp`, then `codex mcp login woopsocial`, then a new session (`codex resume --last` keeps this one) |
| OpenClaw | Install the [WoopSocial skill on ClawHub](https://clawhub.ai/woopsocial/skills/social-media-scheduler), or add the MCP URL to the agent's MCP config |
| Agents without sign-in support | Create an API key at [app.woopsocial.com/api-access](https://app.woopsocial.com/api-access) and use `https://api.woopsocial.com/mcp?api_key=%2A%2A%2A |

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · references/woopsocial.md (reported line 33)May include surrounding context.

md
| Claude (claude.ai, Desktop, Cowork) | Add WoopSocial from the [Claude directory](https://claude.ai/directory/woopsocial), click Connect and sign in |
| Cursor | Install from [cursor.directory](https://cursor.directory/plugins/woopsocial-social-media-scheduler) and sign in |
| Claude Code | `claude mcp add --transport http woopsocial "https://api.woopsocial.com/mcp"`, then restart with `claude --continue`, `/mcp`, select woopsocial, Authenticate |
| Codex | `codex mcp add woopsocial --url https://api.woopsocial.com/mcp`, then `codex mcp login woopsocial`, then a new session (`codex resume --last` keeps this one) |
| OpenClaw | Install the [WoopSocial skill on ClawHub](https://clawhub.ai/woopsocial/skills/social-media-scheduler), or add the MCP URL to the agent's MCP config |
| Agents without sign-in support | Create an API key at [app.woopsocial.com/api-access](https://app.woopsocial.com/api-access) and use `https://api.woopsocial.com/mcp?api_key=%2A%2A%2A |

External Transmission

Medium
Category
Data Exfiltration
Confidence
98% confidence
Finding

The fallback instructs users to place an API credential in the MCP URL query string, where it can be retained in endpoint configuration, diagnostic output, and request logs despite HTTPS. Exposure could allow unauthorized API actions within the key's scope, potentially including access to client media or social posting; the warning against sharing keys does not prevent these incidental leaks.

Content

Scanner excerpt · references/woopsocial.md (reported line 35)May include surrounding context.

md
| Claude Code | `claude mcp add --transport http woopsocial "https://api.woopsocial.com/mcp"`, then restart with `claude --continue`, `/mcp`, select woopsocial, Authenticate |
| Codex | `codex mcp add woopsocial --url https://api.woopsocial.com/mcp`, then `codex mcp login woopsocial`, then a new session (`codex resume --last` keeps this one) |
| OpenClaw | Install the [WoopSocial skill on ClawHub](https://clawhub.ai/woopsocial/skills/social-media-scheduler), or add the MCP URL to the agent's MCP config |
| Agents without sign-in support | Create an API key at [app.woopsocial.com/api-access](https://app.woopsocial.com/api-access) and use `https://api.woopsocial.com/mcp?api_key=%2A%2A%2A |

The API key is a password: never paste it into shared chats or commit it. Never restart the agent
while `make_clips.py` runs.

Tainted flow: 'tmp' from os.environ.get (line 89, credential/environment) → open (file write)

Medium
Category
Data Flow
Confidence
65% confidence
Finding

Data from a source is assigned to a variable that is later passed to a sink, creating a variable-mediated taint flow.

Content

Scanner excerpt · scripts/agencylib.py (reported line 91)May include surrounding context.

python
os.makedirs(os.path.dirname(os.path.abspath(path)), exist_ok=True)
    tmp = path + ".tmp"
    try:
        with open(tmp, "w", encoding="utf-8") as fh:
            json.dump(data, fh, ensure_ascii=False, indent=1)
        os.replace(tmp, path)
    finally:

Sudo/Root Execution

Medium
Category
Privilege Escalation
Confidence
70% confidence
Finding

Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.

Content

Scanner excerpt · scripts/agencylib.py (reported line 160)May include surrounding context.

python
import shutil
    if not (shutil.which("ffmpeg") and shutil.which("ffprobe")):
        die("ffmpeg not found. Mac: brew install ffmpeg. Windows: winget install Gyan.FFmpeg (then open a "
            "new terminal). Linux: sudo apt install ffmpeg")


@functools.lru_cache(maxsize=None)

Sudo/Root Execution

Medium
Category
Privilege Escalation
Confidence
70% confidence
Finding

Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.

Content

Scanner excerpt · scripts/check_setup.py (reported line 34)May include surrounding context.

python
"install ffmpeg. Windows: winget install Gyan.FFmpeg. Linux: your distro's full ffmpeg package")
    else:
        line("missing", "ffmpeg", "Mac: brew install ffmpeg. Windows: winget install Gyan.FFmpeg (then open a "
             "new terminal). Linux: sudo apt install ffmpeg")
        ok = False
    try:
        import PIL

Sudo/Root Execution

Medium
Category
Privilege Escalation
Confidence
70% confidence
Finding

Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.

Content

Scanner excerpt · scripts/transcribe.py (reported line 49)May include surrounding context.

python
def need_ffmpeg():
    if not shutil.which("ffmpeg") or not shutil.which("ffprobe"):
        die("ffmpeg is not installed. Mac: brew install ffmpeg. Windows: winget install ffmpeg. "
                 "Linux: sudo apt install ffmpeg (or your distro's package).")


def duration(path):

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 · scripts/agencylib.py (reported line 166)May include surrounding context.

python
@functools.lru_cache(maxsize=None)
def ffmpeg_has(kind, name):
    """Whether this ffmpeg build has an encoder or filter."""
    r = subprocess.run(["ffmpeg", "-hide_banner", f"-{kind}"], capture_output=True, text=True,
                       encoding="utf-8", errors="replace")
    return bool(re.search(rf"\s{re.escape(name)}\s", r.stdout))

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 · scripts/agencylib.py (reported line 209)May include surrounding context.

python
if not os.path.getsize(path):
        die(f"{name} is an empty file (0 bytes)")
    need_ffmpeg()
    r = subprocess.run(["ffprobe", "-v", "error", "-show_streams", "-show_format", "-of", "json", path],
                       capture_output=True, text=True, encoding="utf-8", errors="replace")
    try:
        d = json.loads(r.stdout or "{}")

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · scripts/approve.py (reported line 493)May include surrounding context.

python
m = check_batch(batch_path, ledger_file)
    appr = load_approvals(m)
    if not appr["pages"]:
        die("no approval page yet: run approve.py on the batch first")
    pg = newest_page(appr)
    if pg.get("fp") != m["fp"]:
        die("the batch changed after the last page was made: make a new page (approve.py) and send that one")

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · scripts/approve.py (reported line 722)May include surrounding context.

python
body = strip_quoted(text)
    appr = load_approvals(m)
    if not appr["pages"]:
        die("no approval page was made for this episode yet")
    named = {int(x) for x in re.findall(r"(?i)\bpage[ \t ]*#?[ \t ]*(\d{1,3})\b", unicodedata.normalize("NFKC", body))}
    if len(named) > 1:
        die(f"the reply names more than one page ({', '.join(map(str, sorted(named)))}); apply one page's answers at a time")

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 · scripts/check_clips.py (reported line 199)May include surrounding context.

python
fd, tmp = tempfile.mkstemp(suffix=".f32")
    os.close(fd)
    try:
        subprocess.run(["ffmpeg", "-v", "error", "-y", "-ss", f"{max(0.0, start):.3f}", "-t", f"{length:.3f}", "-i", path,
                        "-vn", "-ac", "1", "-ar", "16000", "-f", "f32le", tmp], capture_output=True)
        audio = np.fromfile(tmp, dtype=np.float32)
    finally: