Back to skill

Security audit

subclip-tts

Security checks for vulnerabilities and agentic risk

Overview

ClawScan marks this skill suspicious because scanners report conflicting guidance that could send an intended offline script to paid cloud speech without separate approval.

Before installing, consider whether you are comfortable with the reported ambiguity around cloud speech: the offline fallback could send a script to a cloud service and incur credits. The scan is partial because several local path references could not be resolved, so the evidence does not establish that every referenced item was inspected.

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
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (4)

The offline workflow instructs a credit-consuming cloud fallback despite promising offline, no-credit generation.

Medium
Category
Excessive Agency
Confidence
97% confidence
Finding

The offline workflow explicitly applies when the creator wants no credit expenditure or is offline, yet directs cloud generation when local speech is unavailable. This can transmit a script intended to stay local and incur unwanted charges; the later cloud-only-on-request restriction creates ambiguity rather than reliably preventing the fallback.

Content

No source excerpt is available for this finding.

The offline fallback directs the agent to use paid cloud speech despite the creator's no-credit constraint.

Medium
Category
Prompt Injection
Confidence
97% confidence
Finding

The instruction to use cloud speech overrides the no-credit constraint that selects this workflow. Although other passages restrict cloud use, this explicit fallback can cause an agent to treat authorization for offline generation as authorization for a paid external operation.

Content

No source excerpt is available for this finding.

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 · SKILL.md (reported line 41)May include surrounding context.

md
ine speech is not enabled in this build yet, so use `generate_speech` with a cloud voice and say so. Cloning still works.
2. Pick a voice `id` from the list: a built-in voice such as `alba`, or a cloned voice (`clone:...`) the creator made on this computer.
3. Call `generate_local_speech` with the exact script and `voice`. For a language other than English pass `modelId: "subclip-local-supertonic"` and `language` (31 languages, voices `M1` to `M5` and `F1` to `F5`); `get_voice_samples` returns these arguments ready to use. It needs no confirmation and uses no credits. It returns the WAV `filePath`. Do not poll.
4. Tell the creator where the file is and that it was made offline. Use `generate_speech` only when the creator asks for a cloud voice or a different model.

Cloned voices are made from the AI Assets speech picker (Clone a voice offline): 3 to 30 seconds of clear speech from one speaker, with the creator's permission. Never clone a voice for someone who has not given permission.

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 · references/model-selection.md (reported line 11)May include surrounding context.

md
ability and do not invent IDs.
- Speech: choose a `text_to_speech` model and use `list_ai_voices`; “All models” is a filter, not a generation model. Voice IDs belong to their model. Inworld Realtime TTS 2 is for its preset voices, not Subclip saved clones. Gemini separates style instructions from script; ElevenLabs v3 supports expressive delivery, while Flash prioritizes short-latency speech; MiniMax supports its documented interjection and pause syntax. Never copy one model's tags into another without checking. Use supported languages and the live billing unit; characters and tokens are not interchangeable.
- Cloning: use the existing Subclip saved-voice workflow and its stable IDs. Do not send private cloned IDs to Replicate/fal. Keep clone UI/provider descriptions neutral.

Retired choices for new jobs: LTX 2.3 Pro, Wan 2.7, FLUX 1.1 Pro Ultra, Qwen Image 2 Pro and Seedream 5 Pro. Their historical outputs remain usable. A remix needs an active model and new estimate, with confirmati