Back to skill

Security audit

skill-promo-video

Security checks for vulnerabilities and agentic risk

Overview

ClawScan rates this skill suspicious because it found concerns about unrestricted network destinations and shell command handling; Skillspector coverage was partial, so the evidence is incomplete.

Consider the suspicious verdict and the reported risks before installing. In particular, review whether the skill needs access to shell and files, whether any supplied service URL could send job details outside your machine, and whether its scripts could overwrite files or clear a Blender scene. Skillspector scanned all listed components but marked analysis partial because two local references were unresolved, so the available evidence is incomplete.

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
Vulnerability Patterns
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • 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
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (22)

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

The script deletes all objects in the active Blender scene without disclosing that existing scene content will be removed.

Medium
Category
Output Handling
Confidence
90% confidence
Finding

The main workflow unconditionally calls clear_scene(), which selects and deletes every scene object and removes unused data blocks. The module documentation describes rendering and output files but does not warn about clearing an existing scene; there is no confirmation or user-facing log for this operation.

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 · scripts/card_composite.py (reported line 23)May include surrounding context.

python
def probe(p):
    r = subprocess.run(["ffprobe", "-v", "quiet", "-show_entries", "format=duration",
                        "-of", "csv=p=0", p], capture_output=True, text=True)
    return float(r.stdout.strip())

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/compose_3step.py (reported line 17)May include surrounding context.

python
sys.stdout.reconfigure(encoding="utf-8")

def probe_dur(p):
    r = subprocess.run(["ffprobe", "-v", "quiet", "-show_entries", "format=duration",
                        "-of", "csv=p=0", p], capture_output=True, text=True).stdout.strip()
    assert r, f"missing/empty {p}"
    return round(float(r), 3)

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/card_composite.py (reported line 65)May include surrounding context.

python
"setsar=1,fps=30[base];[base][1:v]overlay=%d:0:format=auto" % (W, H, W, H, x0)),
           "-an", "-c:v", "libx264", "-crf", "18", "-preset", "medium",
           "-movflags", "+faststart", a.out]
    r = subprocess.run(cmd, capture_output=True, text=True)
    assert r.returncode == 0, r.stderr[-400:]
    print("COMPOSITE OK:", a.out, "%.2fs" % probe(a.out))

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
90% confidence
Finding

Although the subprocess does not use a shell, the user-controlled ASS filename is interpolated directly into FFmpeg's filtergraph syntax. A crafted filename containing filter delimiters can alter the graph, potentially introducing additional file or network sources accessible to FFmpeg; ordinary punctuation can also break processing.

Content

Scanner excerpt · scripts/compose_3step.py (reported line 90)May include surrounding context.

python
cmd += ["-filter_complex", ";".join(parts), "-map", "[vf]",
            "-c:v", "libx264", "-crf", "19", "-preset", "medium", "-r", str(FPS),
            "-t", f"{total:.2f}", os.path.join(tmpd, "video.mp4")]
    r = subprocess.run(cmd, capture_output=True, text=True)
    if r.returncode != 0:
        print("STEP1 FAIL:", r.stderr[-1200:]); sys.exit(1)
    print("step1 video ok", total)

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/compose_3step.py (reported line 112)May include surrounding context.

python
acmd += ["-filter_complex", ";".join(aparts), "-map", "[aout]",
             "-c:a", "aac", "-b:a", "160k", "-t", f"{total:.2f}", os.path.join(tmpd, "audio.m4a")]
    import os as _os, time as _t
    r = subprocess.run(acmd, capture_output=True, text=True)
    if r.returncode != 0:
        print("STEP2 FAIL:", r.stderr[-1200:]); sys.exit(1)
    print("step2 audio ok")

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/compose_3step.py (reported line 118)May include surrounding context.

python
print("step2 audio ok")

    # ── Step3 合并 ──
    r = subprocess.run(["ffmpeg", "-y", "-i", os.path.join(tmpd, "video.mp4"), "-i", os.path.join(tmpd, "audio.m4a"),
                        "-c:v", "copy", "-c:a", "aac", "-b:a", "160k",
                        "-movflags", "+faststart", a.out], capture_output=True, text=True)
    if r.returncode != 0:

The documented guarantee of a fresh, unique output is contradicted by unconditional overwriting of the requested output path.

Medium
Category
Excessive Agency
Confidence
98% confidence
Finding

The workflow calls for fresh, unique output names, but the final FFmpeg invocation uses -y and does not check whether the destination already exists. An accidental collision or an untrusted output-path selection can silently destroy an existing writable file, although the damage is limited by the executing user's permissions.

Content

No source excerpt is available for this finding.

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
86% confidence
Finding

The request URL derives from an unrestricted --t2i-url value, despite the skill's explicit loopback-only network boundary. If an untrusted task supplies that value, polling can contact a remote host and disclose request metadata and job identifiers; this is an endpoint-validation issue rather than shell injection.

Content

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

python
def http_json(url, payload=None, timeout=20):
    if payload is None:
        r = subprocess.run(["curl", "-s", "-m", str(timeout), url], capture_output=True, text=True)
    else:
        pf = os.path.join(tempfile.mkdtemp(prefix="spv_pl_"), "payload.json")
        json.dump(payload, open(pf, "w"), ensure_ascii=False)

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
91% confidence
Finding

The POST destination accepts an arbitrary --t2i-url without enforcing the advertised local-only restriction. A remote destination would receive the generation graph, including supplied character descriptions, scene text, and the reference-image filename; the shown payload does not upload the reference-image bytes.

Content

Scanner excerpt · scripts/gen_scene_cards.py (reported line 95)May include surrounding context.

python
else:
        pf = os.path.join(tempfile.mkdtemp(prefix="spv_pl_"), "payload.json")
        json.dump(payload, open(pf, "w"), ensure_ascii=False)
        r = subprocess.run(["curl", "-s", "-m", str(timeout), "-X", "POST", "-H",
                            "Content-Type: application/json", "-d", "@" + pf, url],
                           capture_output=True, text=True)
        os.remove(pf)

The documented character-by-character OCR verification is not performed before generated cards are accepted.

Medium
Category
Excessive Agency
Confidence
98% confidence
Finding

L009 explicitly requires character-by-character OCR verification after image generation, but the --wait completion path accepts an image solely because its file size exceeds 100,000 bytes. It then marks the image OK without inspecting its text, allowing incorrect promotional claims or numbers to pass the documented verification gate.

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 · scripts/measure_sync.py (reported line 23)May include surrounding context.

python
def lip_curve(video, fps=24):
    tmp = tempfile.mkdtemp(prefix="sync_")
    subprocess.run(["ffmpeg", "-y", "-v", "quiet", "-i", video, "-vf", "scale=480:-1",
                    "-f", "image2", os.path.join(tmp, "f%04d.png")], check=True)
    fs = sorted(f for f in os.listdir(tmp) if f.endswith(".png"))
    frames = [np.asarray(Image.open(os.path.join(tmp, f)).convert("L"), dtype=np.float32)

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/qc_gates.py (reported line 26)May include surrounding context.

python
def sh(cmd, timeout=60):
    return subprocess.run(cmd, capture_output=True, text=True, timeout=timeout)


def dur(p):

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
99% confidence
Finding

The ASR gate interpolates video and requirement values directly into a command subsequently interpreted by bash. If an attacker influences either value, shell metacharacters or command substitutions can execute commands with the agent's permissions; intentionally supporting a trusted command template does not make interpolated data safe.

Content

Scanner excerpt · scripts/qc_gates.py (reported line 132)May include surrounding context.

python
def cmd_asr(a):
    assert a.gate_cmd, "--gate-cmd 必填(外部 FunASR 闸门命令模板,{media} {require} 占位)"
    cmd = a.gate_cmd.format(media=a.video, require=a.require or "")
    r = subprocess.run(["bash", "-c", cmd], capture_output=True, text=True, timeout=600)
    print(r.stdout[-1500:])
    if r.returncode != 0:
        print("ASR GATE FAIL"); sys.exit(1)

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/style_transfer.py (reported line 69)May include surrounding context.

python
import tempfile
    pf = os.path.join(tempfile.mkdtemp(prefix="spv_style_"), "payload.json")
    json.dump({"prompt": graph}, open(pf, "w"), ensure_ascii=False)
    r = subprocess.run(["curl", "-s", "-m", "30", "-X", "POST", "-H", "Content-Type: application/json",
                        "-d", "@" + pf, a.edit_url + "/prompt"], capture_output=True, text=True)
    resp = json.loads(r.stdout)
    assert "prompt_id" in resp, "SUBMIT FAIL: " + r.stdout[:300]

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/style_transfer.py (reported line 77)May include surrounding context.

python
t0 = time.time()
    while (time.time() - t0) < a.timeout_min * 60:
        time.sleep(10)
        h = json.loads(subprocess.run(["curl", "-s", a.edit_url + "/history/" + pid],
                                      capture_output=True, text=True).stdout or "{}")
        node = h.get(pid)
        if not node:

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/talking_segment.py (reported line 42)May include surrounding context.

python
def sh(cmd, timeout=30):
    return subprocess.run(cmd, capture_output=True, text=True, timeout=timeout)


def audio_seconds(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/tts_batch.py (reported line 24)May include surrounding context.

python
for seg in segs:
        wav = os.path.join(a.out_dir, seg["id"] + ".wav")
        txt = seg["tts"].replace("?", "。").replace("?", "。")  # pitfalls #17:问号静默失败
        r = subprocess.run(["bash", a.medtts, txt, wav],
                           capture_output=True, text=True, timeout=300)
        # pitfalls #2:rc=0 不可信,验文件+mtime+size
        if (not os.path.exists(wav)) or os.path.getmtime(wav) < t0 or os.path.getsize(wav) < 1000:

Delivery filenames are required to be in Chinese without user language choice.

Low
Category
Output Handling
Confidence
97% confidence
Finding

The instruction explicitly requires clear Chinese names for files delivered through Feishu. Although avoiding internal version numbers is justified, requiring Chinese is not justified by that rationale, and no user language choice or opt-in is provided.

Content

No source excerpt is available for this finding.

Video frames are written to a persistent temporary directory without disclosure.

Low
Category
Output Handling
Confidence
95% confidence
Finding

The script extracts video frames into an mkdtemp directory and never removes them, contrary to the workflow's disposable-temporary-data description. This unnecessarily retains potentially private imagery and consumes disk space; the private directory permissions limit exposure, so it is not evidence of public disclosure.

Content

No source excerpt is available for this finding.

The S2V path does not use the promised native 720p resolution; it defaults to the H3 dimensions instead.

Low
Category
Excessive Agency
Confidence
99% confidence
Finding

The manifest promises native 720p S2V generation, and L013 explicitly identifies native resolution as necessary for lip quality. However, the S2V graph uses a.width and a.height, whose defaults are 1344×768 (L264–L265), while the dedicated 1280×720 S2V arguments (L270–L271) are never used; even supplying those S2V-specific flags does not change the generated resolution.

Content

No source excerpt is available for this finding.