SenseAudio-Let-Claw-Talk

• Outcome: Pass
• Latest audit: May 28, 2026 · 5:57 AM UTC
• Version: 1.0.0

Overview

This skill does what it says: it runs a macOS always-listening voice assistant using SenseAudio, with expected microphone, network, local state, and optional voiceprint features.

Install only if you are comfortable with a long-running microphone assistant that sends captured speech to SenseAudio for transcription and may store preferences and optional voiceprint data locally. Review SenseAudio’s data handling terms, keep WeSpeaker disabled unless you need speaker filtering, and know how to stop the Terminal session or clear stored profiles.

SkillSpector

Vulnerability Patterns

• Data Exfiltration: External Transmission, Env Variable Harvesting, File System Enumeration
• Behavioral AST: exec() Call, eval() Call, Dynamic Import
• MCP Least Privilege: Underdeclared Capability, Wildcard Permission, Missing Permission Declaration
• MCP Tool Poisoning: Hidden Instructions, Unicode Deception, Parameter Description Injection
• Prompt Injection: Instruction Override, Hidden Instructions, Exfiltration Commands

Findings (10)

subprocess module call

Medium

• Category: Dangerous Code Execution

Content

"--message",
        prompt,
    ]
    completed = subprocess.run(command, capture_output=True, text=True, check=False)
    if completed.returncode != 0:
        return None
    payload = extract_json_object(extract_agent_reply(completed.stdout) or completed.stdout)

• Confidence: 89% confidence
• Finding: This function sends user-derived text into an external agent process and then trusts the agent's JSON output to drive voice-routing decisions. Because the code treats another model/tool's output as authoritative control data, prompt injection or hostile agent behavior can manipulate local assistant state and trigger unintended actions.

subprocess module call

Medium

• Category: Dangerous Code Execution

Content

last_error = "audioclaw agent failed"

    for attempt in range(1, max(1, args.agent_retries) + 1):
        completed = subprocess.run(command, capture_output=True, text=True, check=False)
        if completed.returncode == 0:
            reply = extract_agent_reply(completed.stdout)
            if reply:

• Confidence: 92% confidence
• Finding: This executes an external agent binary and uses its output as the assistant's trusted reply path in a continuously listening workflow. In this context, a compromised or prompt-injected agent can steer spoken output, manipulate user decisions, or influence subsequent assistant behavior without any trust boundary enforcement.

Lp3

Medium

• Category: MCP Least Privilege
• Confidence: 94% confidence
• Finding: The skill declares no explicit permissions, yet its documented behavior clearly uses sensitive capabilities: shell execution, network access, reading credentials, and reading/writing state files. This creates a permission-transparency gap where a user or host system may not understand the real trust boundary before invoking a long-lived microphone assistant.

Description-Behavior Mismatch

Medium

• Confidence: 91% confidence
• Finding: The skill description frames WeSpeaker as optional, but the script exposes conversational commands to enroll voice samples, enable speaker filtering, and operate a background service as first-class features. That mismatch matters because users may not expect biometric processing and local access control behavior to be available through casual speech commands in a continuously listening assistant.

Missing User Warnings

Medium

• Confidence: 97% confidence
• Finding: This skill is designed for always-on microphone listening and sends user speech to external SenseAudio services for ASR/TTS, but the description does not prominently warn users about continuous capture, remote processing, or retention/privacy implications. In context, this is more dangerous because the skill is specifically marketed as a persistent desktop voice assistant, increasing the chance of capturing unintended or sensitive conversations.

Missing User Warnings

Medium

• Confidence: 91% confidence
• Finding: The documentation exposes voice cloning and voiceprint enrollment actions without any warning about biometric collection, consent, retention, or impersonation risks. In a continuously listening local voice assistant, this increases the chance that users enable sensitive biometric features without understanding that cloned voices and enrolled voiceprints are privacy-sensitive and can be abused for tracking, spoofing, or unauthorized profile creation.

Missing User Warnings

Medium

• Confidence: 89% confidence
• Finding: The code uploads raw audio content to a third-party ASR service, but this file contains no consent, disclosure, or gating mechanism before transmitting potentially sensitive microphone recordings. In the context of a continuous-listening desktop assistant, that raises the privacy risk substantially because captured speech may include confidential or unintended background conversations.

Missing User Warnings

Medium

• Confidence: 88% confidence
• Finding: The /clear_profile endpoint deletes stored profile files immediately with no authentication, authorization, or confirmation step. Although the service binds to 127.0.0.1 by default, any local process can invoke it, allowing accidental or malicious deletion of enrolled voice profiles and denial of service for speaker verification.

Missing User Warnings

Medium

• Confidence: 90% confidence
• Finding: The script records microphone input directly to a caller-specified file path and emits only machine-readable JSON after capture, with no explicit user-facing disclosure at runtime that audio is being saved to disk. In the context of a skill explicitly designed for continuous listening on macOS, this increases privacy risk because users may not understand when recording starts, where audio is stored, or how long it persists, enabling accidental collection of sensitive speech.

Missing User Warnings

Medium

• Confidence: 96% confidence
• Finding: The enrollment flow records and submits the user's voice for WeSpeaker profile creation, which is biometric data processing. The code provides convenience prompts but no meaningful disclosure about storage location, retention, transmission to the background service, or the sensitivity of voiceprints, creating a real privacy and consent problem.

Static analysis

No suspicious patterns detected.