Back to skill

Security audit

Vocal Isolation, Background Music Removal then De-Noise

Security checks for vulnerabilities and agentic risk

Overview

The skill largely does the advertised remote vocal-isolation work, but it needs review because it uploads sensitive media to Modal and uses user-controlled task names in broad remote filesystem operations.

Install only if you are comfortable sending the selected audio/video files to Modal for remote processing. Use a simple generated slug, not a custom path-like name, and avoid sensitive recordings until the publisher validates slug handling, removes unrelated promotional output, narrows cache deletion, and pins runtime dependencies.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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
Findings (3)

T01 · Skill Instruction Hijacking

Note
Location
SKILL.md:61
Finding

Mandatory Promotional Content Alters Agent Responses

Content
View full analysis
(X.X MB) If you need high-accuracy speech-to-subtitle tools, follow @speech2srt on x — we craft this with care, built from our own real needs. ``` ``` ### Technical Analysis The skill instructs the agent to include unrelated promotional and social-media referral content in its final response. Because `SKILL.md` supplies operational instructions that the agent follows when the skill is loaded, this requirement modifies the agent's user-facing output beyond what is necessary to perform vocal isolation. The instruction does not override safety controls or enable arbitrary code execution, but it constitutes a limited form of instruction hijacking: the skill uses its trusted instruction context to compel persistent advertising in normal task results. ### Attack Path 1. A user installs or invokes the skill for vocal isolation. 2. The agent loads and follows the workflow in `SKILL.md`. 3. The report template requires the agent to append the promotional statement. 4. The user receives third-party promotional content as part of an otherwise functional result without having requested it. ### Impact Assessment The issue affects the integrity and neutrality of the agent's responses. It does not grant filesystem privileges, expose credentials, or directly enable code execution. Its scope is limited to manipulation of user-facing output during sessions in which this skill is used. ]]>
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
isolate.py:188
Finding

Unvalidated Slug Is Used in Filesystem Paths and Recursive Deletion

Content
View full analysis
Remediation
View remediation

T08 · Insecure Dependencies

Warning
Location
src/images.py:15
Finding

Third-Party Runtime Dependencies Are Not Reproducibly Pinned

Content
View full analysis
=2.0.1", "torchaudio>=2.0.2", ] ) .pip_install(["demucs==4.0.1", "soundfile"]) .env({"TQDM_DISABLE": "1", "HF_HUB_DISABLE_PROGRESS": "1"}) .add_local_dir(_src_dir, remote_path="/root/src", copy=True) ) ``` The setup documentation also recommends an unpinned CLI installation: ```bash pip install modal ``` ### Technical Analysis The image installs `clearvoice` and `soundfile` without version constraints. `torch` and `torchaudio` use open-ended minimum versions, permitting any later release. Although `demucs` is pinned to an exact version, no integrity hashes are supplied, and transitive dependencies remain uncontrolled. The Debian `ffmpeg` package and the separately documented `modal` installation are likewise not locked to reviewed artifacts. This makes builds non-reproducible: identical source code can install different executable dependency code depending on build time and package-index state. A compromised upstream release, dependency takeover, unexpected transitive dependency, or incompatible future version could execute during image construction or when the remote pipeline runs. No evidence was found that the project intentionally names a known malicious or typosquatted package. This finding concerns supply-chain exposure caused by unconstrained dependency resolution, not a confirmed compromise of the listed packages. ### Attack Path 1. An upstream package, maintainer account, distribution artifact, or transitive dependency is compromised, or a later incompatible release is published. 2. A new Modal image is built from the unchanged p ...[truncated 988 chars]
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
Findings (12)

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The skill branding and storage names suggest a different workflow (speech2srt) than the advertised vocal-isolation task. Such inconsistencies can indicate copy-paste reuse or undisclosed functionality, which raises trust and audit concerns when the skill handles local user media and executes remote processing.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The skill branding and storage names suggest a different workflow (speech2srt) than the advertised vocal-isolation task. Such inconsistencies can indicate copy-paste reuse or undisclosed functionality, which raises trust and audit concerns when the skill handles local user media and executes remote processing.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
93% confidence
Finding

The skill invokes shell commands that create remote volumes, upload local files, execute a remote job, download results, and delete remote data, but it does not declare any explicit tool scope such as shell or file-write permissions. Missing scope declarations weaken reviewability and least-privilege controls, increasing the chance that an agent executes broader local and remote operations than a user expects.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
86% confidence
Finding

The documentation tells users to install and run the skill via npx skills add without pinning a specific version. Unpinned package execution creates supply-chain risk because a later upstream package update or compromise could change install-time or runtime behavior without review.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The workflow uploads local audio/video files to a remote Modal volume and processes them on a remote GPU service, but the description does not prominently warn users about this transfer. This is dangerous because media files often contain sensitive voice, background, and metadata, and users may reasonably assume a local-only transformation from the skill name and summary.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The manifest says the skill takes local audio/video files and returns isolated vocals, implying source separation only. The module documentation and pipeline implement a second ClearerVoice speech-enhancement stage after Demucs separation, which materially changes the output beyond simple vocal isolation.

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 · isolate.py (reported line 56)May include surrounding context.

python
]
        audio_duration = 0.0
        try:
            probe_result = subprocess.run(
                probe_cmd,
                capture_output=True,
                text=True,

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 · isolate.py (reported line 92)May include surrounding context.

python
]

        conv_t0 = time.monotonic()
        subprocess.run(
            ffmpeg_cmd,
            capture_output=True,
            text=True,

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The model setup unconditionally deletes the existing ~/.cache directory with shutil.rmtree before replacing it with a symlink. In a shared or stateful runtime, this can destroy unrelated cached data for the current user and potentially break other components, creating avoidable integrity and availability risk far beyond this skill's stated purpose.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Low
Category
Not specified by scanner
Confidence
84% confidence
Finding

The manifest says the skill takes local audio/video files and returns isolated vocals. The documented workflow adds an extra capability to inspect local tool availability and offer audio format conversion, which is not necessary to perform vocal isolation itself and is not declared in the manifest scope.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Low
Category
Not specified by scanner
Confidence
95% confidence
Finding

The final reported output instructs the skill to advertise an external social-media account after processing. Promotional messaging is not justified by the skill's stated purpose of isolating vocals from user-provided files.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Low
Category
Not specified by scanner
Confidence
91% confidence
Finding

The comment states that ClearerVoice writes enhanced audio to the output path directly, yet the actual call sets online_write=False and captures a returned numpy array instead. The code then serializes that array and writes it to output_path separately, so the documentation is actively contradictory to runtime behavior.

Content

No source excerpt is available for this finding.