Back to skill

Security audit

Local Voice Agent

Security checks for vulnerabilities and agentic risk

Overview

This voice assistant skill is mostly purpose-aligned, but it overstates its privacy and completeness and includes unsafe configuration, network, and dependency practices that should be reviewed before use.

Install only if you are comfortable reviewing and hardening it first: remove the shell `eval`, keep the TTS service bound to localhost, avoid remote HTTP TTS unless trusted, pin downloaded dependencies/models, and treat microphone recordings, transcriptions, and cached speech as sensitive local data.

Vulnerability Patterns
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
Findings (4)

T09 · Insecure Skill Coding Practices

Error
Location
bin/voice-agent.sh:106
Finding

Configuration-Based Arbitrary Command Execution Through eval

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
bin/voice-agent.sh:162
Finding

Predictable Audio Files in an Unverified Shared Temporary Directory

Content
View full analysis
/dev/null || \ ffmpeg -f alsa -i default -t "$DURATION" -y "$RECORDING_FILE" 2>/dev/null ``` ### Technical Analysis The script uses a fixed directory beneath globally shared `/tmp` and does not: - Create it atomically. - Verify its owner or permissions. - Set a restrictive `umask`. - Use unpredictable recording names. - Verify that the destination is not a symbolic link. The filename contains only the current Unix timestamp in seconds, making it readily predictable. The `-y` option tells FFmpeg to overwrite the selected output without confirmation. On a multi-user system, an attacker may pre-create `/tmp/voice-agent` before the victim or race to prepare a predictable destination. Depending on directory ownership and filesystem protections, a symbolic link or attacker-controlled file can redirect or interfere with output creation. ### Attack Path 1. The attacker predicts when interactive recording will begin. 2. If the fixed directory does not yet exist, the attacker creates `/tmp/voice-agent` with attacker-controlled permissions. Alternatively, the attacker targets an installation where that directory is improperly shared. 3. The attacker predicts a path such as `/tmp/voice-agent/recording_1750000000.wav`. 4. The attacker places a symbolic link or conflicting file at that location. 5. The victim st ...[truncated 963 chars]
Remediation
View remediation

T03 · Remote Payload Retrieval and Execution

Warning
Location
install.sh:50
Finding

Mutable Third-Party Source Is Downloaded, Built, and Executed Without Integrity Pinning

Content
View full analysis
Remediation
View remediation
" git clone https://github.com/ggerganov/whisper.cpp "$HOME/.local/whisper.cpp" git -C "$HOME/.local/whisper.cpp" checkout --detach "$WHISPER_COMMIT" test "$(git -C "$HOME/.local/whisper.cpp" rev-parse HEAD)" = "$WHISPER_COMMIT" ``` - Verify signed tags or release signatures where available. - Download models from a fixed versioned URL and validate a documented SHA-256 digest before use. - Pin Python dependencies to reviewed versions. - Use a requirements file with hashes, for example `pip install --require-hashes -r requirements.txt`. - Install Python dependencies in an isolated virtual environment. - Record the reviewed dependency versions in the project documentation and update them through an explicit review process. - Preserve the current user-consent prompt, but do not treat consent as a replacement for integrity verification. ]]>

T09 · Insecure Skill Coding Practices

Warning
Location
SKILL.md:67
Finding

Documentation Exposes the Unauthenticated TTS Service on All Network Interfaces

Content
View full analysis
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Rogue AgentSelf-Modification, Session Persistence
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
Findings (43)

Credential Access

High
Category
Privilege Escalation
Confidence
60% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · .gitignore (reported line 25)May include surrounding context.

text
logs/

# Environment
.env
.venv/
venv/

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

The description significantly overstates what this code does. The supplied script only converts provided text into speech audio and optionally plays it back. It does not capture microphone input, transcribe speech, perform assistant logic, or provide end-to-end voice-to-voice interaction. Additionally, while the description emphasizes fully local/offline operation, the script depends on a configurable TTS server URL and a Python wrapper that may make HTTP requests, so this chunk does not itself prove exclusively local processing. Therefore the declared description does not accurately represent the actual behavior of this code chunk.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
93% confidence
Finding

The description overstates completeness and functionality. The code does implement substantial voice pipeline behavior: dependency checks, microphone/audio-file input, transcription via a local STT script, TTS generation via a local helper, playback, and interactive looping. However, the central 'process_with_ai' function is explicitly a placeholder that only echoes canned text and says actual OpenClaw integration must be implemented. That makes the declared purpose of a complete OpenClaw voice assistant materially inaccurate. Additionally, while the setup appears intended for local use, it contacts a TTS server over HTTP and merely defaults to localhost based on config; the code chunk alone does not fully substantiate the strict '100% local processing' claim. The extra support for text input and audio-file processing is not harmful, but it is broader than the voice-only framing. Overall, this is a meaningful description-versus-behavior mismatch because the primary promised capability is not actually present.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The declared description presents a full offline voice-to-voice AI assistant with both STT and TTS capabilities and broader assistant use cases. The supplied code chunk only implements a command-line speech-to-text/transcription entrypoint. It checks for python3, PyYAML, and ffmpeg, parses model/language flags, and calls a local stt.py script on an audio file. That is consistent with a local transcription component, but it does not demonstrate the broader declared primary purpose of a complete voice-to-voice assistant, nor any text-to-speech, assistant logic, or voice cloning. This is therefore a material description-to-behavior mismatch rather than a mere partial implementation detail.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

The declared description describes a full offline voice-to-voice assistant, implying both speech input and spoken output plus assistant behavior. The supplied code only implements STT/transcription via Whisper.cpp and microphone capture via ffmpeg. It does operate locally and does not use cloud APIs, which is consistent with part of the description, but the primary scope of the code chunk is much narrower than the declared capability set. This is a material description-to-behavior mismatch rather than a minor omission.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The manifest describes a complete voice-to-voice AI assistant for OpenClaw, implying real end-to-end assistant behavior. However, the core AI step is stubbed out with placeholder output instead of invoking OpenClaw, so the implemented behavior materially falls short of the stated capability.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The configuration text explicitly labels the TTS endpoint as a 'VPS instance' while the skill metadata claims '100% local processing, no cloud APIs.' Even though the current URL is localhost, this contradiction is security-relevant because it can mislead users about where audio-derived data may be sent, causing unintended exposure of speech content or generated outputs if deployed with a non-local endpoint.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The README makes strong privacy/security claims such as '100% local processing' and 'No Cloud APIs' while also documenting a configurable HTTP TTS endpoint that can be pointed at a non-local server. This can mislead users into believing voice data never leaves the machine, increasing the risk of accidental transmission of sensitive audio or text to remote infrastructure.

Content

No source excerpt is available for this finding.

Sudo/Root Execution

Medium
Category
Privilege Escalation
Confidence
70% confidence
Finding

Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.

Content

Scanner excerpt · README.md (reported line 52)May include surrounding context.

export POCKET_TTS_URL="http://localhost:5000"

FFmpeg (Audio Conversion)

sudo apt-get install -y ffmpeg

text

### 2. Install Skill

Sudo/Root Execution

Medium
Category
Privilege Escalation
Confidence
70% confidence
Finding

Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.

Content

Scanner excerpt · README.md (reported line 192)May include surrounding context.

export POCKET_TTS_URL="http://localhost:5000"

FFmpeg (Audio Conversion)

sudo apt-get install -y ffmpeg

text

### 2. Install Skill

Sudo/Root Execution

Medium
Category
Privilege Escalation
Confidence
70% confidence
Finding

Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.

Content

Scanner excerpt · SKILL.md (reported line 80)May include surrounding context.

export POCKET_TTS_URL="http://localhost:5000"

FFmpeg (Audio Conversion)

sudo apt-get install -y ffmpeg

text

### 2. Install Skill

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
92% confidence
Finding

The skill advertises and documents capabilities that involve shell execution, filesystem access, and network communication, but it does not declare any explicit tool scope such as permissions or allowed-tools. That creates an authorization ambiguity where an agent may grant broader access than users expect, increasing the risk of unintended command execution, file access, or service communication.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The documentation instructs users to send voice recordings and receive generated responses through an HTTP service without warning that this data may contain sensitive personal or operational information. Even when the default endpoint is localhost, a local or misconfigured service can log, retain, proxy, or expose recordings and transcriptions, leading to privacy and confidentiality risks.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
85% confidence
Finding

The description emphasizes fully local processing, but the code treats TTS as an HTTP service identified by a configurable URL and probes it with curl. While the default points to localhost, the implementation permits non-local endpoints, so the code does not enforce the manifest's strict local-only claim.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

This code records from the system microphone and writes the captured audio to a temporary WAV file, which is a privacy-relevant operation. Although it prints that recording is starting, it does not clearly warn the user that spoken audio will be captured and stored in /tmp before transcription.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

Caching STT/TTS artifacts creates retention of audio-derived data on disk, which can expose sensitive spoken content, transcriptions, or synthesized outputs to other local users, backups, or forensic recovery. In a voice-assistant context, this is more dangerous because users may speak passwords, personal data, health information, or private requests without realizing those artifacts persist.

Content

No source excerpt is available for this finding.

Session Persistence

Medium
Category
Rogue Agent
Confidence
60% confidence
Finding

Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.

Content

Scanner excerpt · install.sh (reported line 35)May include surrounding context.

sh
SKILLS_DIR="$HOME/.openclaw/workspace/skills/voice-agent"
echo -e "${YELLOW}📦 Installing to: $SKILLS_DIR${NC}"

mkdir -p "$SKILLS_DIR"
cp -r "$VOICE_AGENT_DIR"/* "$SKILLS_DIR/"

# Make scripts executable

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The manifest emphasizes fully local processing with no cloud APIs, but this installer clones Whisper.cpp from GitHub and downloads a model over the network. While the runtime voice processing may remain local, the skill as delivered does perform remote network operations during installation, which materially exceeds the plain-language '100% local' claim.

Content

No source excerpt is available for this finding.

Sudo/Root Execution

Medium
Category
Privilege Escalation
Confidence
70% confidence
Finding

Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.

Content

Scanner excerpt · bin/voice-agent.sh (reported line 101)May include surrounding context.

sh
if ! command -v ffmpeg &> /dev/null; then
    echo ""
    echo -e "${YELLOW}⚠️  FFmpeg not installed${NC}"
    echo -e "${BLUE}💡 Install with: sudo apt-get install -y ffmpeg${NC}"
else
    echo -e "${GREEN}✅ FFmpeg already installed${NC}"
fi

Sudo/Root Execution

Medium
Category
Privilege Escalation
Confidence
70% confidence
Finding

Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.

Content

Scanner excerpt · bin/voice-to-text.sh (reported line 87)May include surrounding context.

sh
if ! command -v ffmpeg &> /dev/null; then
    echo ""
    echo -e "${YELLOW}⚠️  FFmpeg not installed${NC}"
    echo -e "${BLUE}💡 Install with: sudo apt-get install -y ffmpeg${NC}"
else
    echo -e "${GREEN}✅ FFmpeg already installed${NC}"
fi

Sudo/Root Execution

Medium
Category
Privilege Escalation
Confidence
70% confidence
Finding

Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.

Content

Scanner excerpt · install.sh (reported line 69)May include surrounding context.

sh
if ! command -v ffmpeg &> /dev/null; then
    echo ""
    echo -e "${YELLOW}⚠️  FFmpeg not installed${NC}"
    echo -e "${BLUE}💡 Install with: sudo apt-get install -y ffmpeg${NC}"
else
    echo -e "${GREEN}✅ FFmpeg already installed${NC}"
fi

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 · lib/stt.py (reported line 57)May include surrounding context.

python
return path

    # Try system path
    if subprocess.run(['which', 'whisper-cli'], capture_output=True).returncode == 0:
        return 'whisper-cli'

    raise FileNotFoundError(

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 · lib/stt.py (reported line 77)May include surrounding context.

python
wav_path = tmp.name

    try:
        subprocess.run([
            'ffmpeg', '-i', audio_path,
            '-ar', '16000',  # 16kHz sample rate
            '-ac', '1',      # Mono

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 · lib/stt.py (reported line 194)May include surrounding context.

python
'--no-timestamps'  # Clean output without timestamps
        ]

        result = subprocess.run(cmd, capture_output=True, text=True, check=True)

        # Extract transcription (last line after processing info)
        output_lines = result.stdout.strip().split('\n')

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The function records from the microphone immediately and proceeds to transcription with only a console status message, not a clear consent or privacy warning. In a voice-agent skill, microphone capture is sensitive data collection; if invoked unexpectedly by another component, it can capture private speech without informed user consent.

Content

No source excerpt is available for this finding.