Back to skill

Security audit

digital-human-spokesperson

Security checks for vulnerabilities and agentic risk

Overview

This skill is a disclosed media-generation pipeline with expected local tooling and optional cloud rendering, but users should understand the privacy implications before using cloud engines.

Install only if you are comfortable with a media-generation skill that runs ffmpeg/SadTalker locally, may require large third-party dependencies, and can send avatar images and voice/audio to D-ID or HeyGen when cloud engines are chosen. Use the local SadTalker or fallback path for sensitive media, and review any public URL hosting before using HeyGen.

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
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • 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 (17)

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The declared description promises an end-to-end bilingual video generator from English copy: automatic translation, English and Arabic voice generation, subtitle generation, avatar creation, and composition into one single video with English then Arabic speech. The actual code only takes an existing image and an existing audio file and turns them into one talking-head video clip. It supports several rendering backends (SadTalker, D-ID, HeyGen, or a non-lip-sync ffmpeg fallback), but none of the bilingual, translation, TTS, subtitle, or multi-segment composition behavior exists in this code. This is a material mismatch in primary purpose and capabilities, not just an implementation detail.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The declared description promises an end-to-end bilingual digital-human video generator: avatar creation, English then Arabic narration, subtitles, optional translation, and final single video composition. The supplied code does only one narrow supporting step: synthesize one audio file from one input text using edge-tts, then save metadata including duration. While this could be part of a larger pipeline, this chunk by itself does not implement the skill's primary claimed behavior and lacks several central capabilities (video/avatar/subtitles/translation/bilingual composition). Therefore the description does not accurately represent what this code chunk actually does.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
89% confidence
Finding

The skill instructs use of shell commands, environment variables, file I/O, networked services, and external tooling, but declares no explicit tool scope or permissions. This creates an authorization and review gap: a platform or agent may grant broader capabilities than users or reviewers expect, increasing the chance of unsafe command execution, secret exposure, or uncontrolled external access.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The overview states that the presenter delivers the script in English first and then automatically switches to Arabic, making a specific language workflow the default behavior. While the skill is bilingual by design, the document does not consistently frame Arabic output as an explicit user choice, which can conflict with language/locale policy expectations requiring opt-in or clear justification.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

Overly broad trigger phrases can cause unintended invocation of a skill that performs shell commands, file writes, network access, and media generation. In a tool-using agent environment, accidental activation may waste resources, overwrite files, call external services, or process user content without clear intent.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The cloud lip-sync sections instruct users to upload or transmit local image/audio assets to third-party services and, for HeyGen, to host them on public URLs, but they do not include any privacy, consent, retention, or data-handling warning. Because these assets may contain biometric likeness, voice data, or sensitive content, users could unknowingly expose personal media to external processors or to public internet access.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The markdown says the composition step applies mandatory AI-content labels using Chinese text such as "AI 生成" and "本视频由 AI 生成" by default. This forces a specific language/locale in the output, and the file does not state that the skill is limited to a China-specific compliance context or offer an opt-in choice.

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/compose_bilingual.py (reported line 103)May include surrounding context.

python
ffprobe = _find_ffprobe(ffmpeg)
    if ffprobe:
        try:
            out = subprocess.check_output(
                [
                    ffprobe,
                    "-v",

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_bilingual.py (reported line 119)May include surrounding context.

python
return float(out.strip())
        except (subprocess.CalledProcessError, ValueError):
            pass
    proc = subprocess.run(
        [ffmpeg, "-i", path], capture_output=True, text=True, errors="replace"
    )
    m = re.search(r"Duration:\s*(\d+):(\d+):(\d+(?:\.\d+)?)", proc.stderr)

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
87% confidence
Finding

This helper executes arbitrary command arrays assembled from user-controlled parameters, including ffmpeg filter strings that embed text, font paths, and subtitle-related values. Although shell injection is avoided by passing a list to subprocess.run, ffmpeg itself parses these arguments and can be abused through untrusted executable selection (--ffmpeg) or maliciously crafted media/filter inputs, making this a real command-execution boundary with tool-level injection and local file exposure risk.

Content

Scanner excerpt · scripts/compose_bilingual.py (reported line 161)May include surrounding context.

python
def run(cmd, cwd=None):
    print("[cmd]", " ".join(cmd))
    subprocess.run(cmd, check=True, cwd=cwd)


def _font_part(fontfile: str) -> str:

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The D-ID execution path uploads the local avatar image and audio to a third-party cloud service, but the script gives no explicit runtime consent prompt or user-facing privacy warning when that transfer occurs. In a skill that processes potentially sensitive user-provided media, silent transmission to an external processor can cause unintended data disclosure and compliance issues even if the functionality is expected.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
86% confidence
Finding

This request sends base64-encoded local image and audio content to D-ID's external API endpoint. In this skill context, external transmission is functionally intended, but it is still security-relevant because user media may contain sensitive personal or biometric information and is exfiltrated outside the local environment.

Content

Scanner excerpt · scripts/talking_head.py (reported line 147)May include surrounding context.

python
"config": {"fluent": True, "result_format": "mp4"},
    }
    req = urllib.request.Request(
        "https://api.d-id.com/talks",
        data=json.dumps(payload).encode(),
        headers=headers,
        method="POST",

External Transmission

Medium
Category
Data Exfiltration
Confidence
60% 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 · scripts/talking_head.py (reported line 158)May include surrounding context.

python
print(f"[talking_head] D-ID talk {talk_id} created, waiting...")
    while True:
        req2 = urllib.request.Request(
            f"https://api.d-id.com/talks/{talk_id}", headers=headers
        )
        with urllib.request.urlopen(req2, timeout=60) as resp:
            st = json.load(resp)

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The HeyGen path requires publicly hosted media URLs and then sends those references to a third-party service, yet the script only mentions this in an error/help message rather than issuing a clear execution-time warning. This increases the risk that users expose private avatar/audio data on public URLs or submit sensitive media to external processing without informed consent.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
88% confidence
Finding

This request submits externally hosted image/audio references to HeyGen, causing third-party processing of user media. In this skill, that is expected behavior, but it remains a true privacy/security concern because the media must be publicly reachable and may expose sensitive content to both the hosting layer and the cloud vendor.

Content

Scanner excerpt · scripts/talking_head.py (reported line 193)May include surrounding context.

python
"dimension": {"width": 1280, "height": 720},
    }
    req = urllib.request.Request(
        "https://api.heygen.com/v2/video/talking_photo",
        data=json.dumps(payload).encode(),
        headers=headers,
        method="POST",

External Transmission

Medium
Category
Data Exfiltration
Confidence
60% 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 · scripts/talking_head.py (reported line 204)May include surrounding context.

python
print(f"[talking_head] HeyGen video {video_id} created, waiting...")
    while True:
        req2 = urllib.request.Request(
            f"https://api.heygen.com/v1/video_status.get?video_id={video_id}",
            headers=headers,
        )
        with urllib.request.urlopen(req2, timeout=60) as resp:

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_edge.py (reported line 41)May include surrounding context.

python
if not ffprobe:
        return None
    try:
        out = subprocess.check_output(
            [
                ffprobe,
                "-v",