Back to skill

Security audit

Nex Voice

Security checks for vulnerabilities and agentic risk

Overview

This skill mostly does what it says, but it needs Review because optional LLM mode can send full transcripts and an API key to a configurable endpoint while storing the key locally in plaintext.

Install only if you are comfortable with recordings, transcripts, and action items being stored locally under ~/.nex-voice. Avoid --use-llm for confidential audio unless you trust the exact configured provider endpoint, and prefer environment variables or a credential manager over saving API keys with nex-voice config set-api-key. Review file permissions on ~/.nex-voice/config.json and use an isolated Python environment with reviewed dependency versions.

Vulnerability Patterns
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • 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
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
Findings (3)

T09 · Insecure Skill Coding Practices

Warning
Location
lib/config.py:112
Finding

LLM API Key Stored in a Plaintext Configuration File Without Enforced Permissions

Content
View full analysis
None: """Save configuration to file""" with open(CONFIG_FILE, 'w') as f: json.dump(self.config_data, f, indent=2) ``` ```python def set_api_key(self, key: str) -> None: """Set API key""" self.set("api_key", key) ``` ### Technical Analysis The `set_api_key()` method places the LLM API key directly in the general configuration dictionary, after which `save()` serializes that dictionary as plaintext JSON in `~/.nex-voice/config.json`. The implementation does not: - Use an operating-system credential manager. - Encrypt the stored credential. - Create the configuration file with mode `0600`. - Enforce mode `0700` on the containing data directory. - Correct permissions on an existing configuration file. - Use an atomic, securely permissioned file replacement operation. The resulting permissions depend on the user's current `umask` and pre-existing filesystem state. On a multi-user system or a machine containing untrusted local processes, the API key may consequently be readable outside its intended security boundary. ### Attack Path 1. A user runs `nex-voice config set-api-key`. 2. The supplied credential is stored under the `api_key` field in `config.config_data`. 3. `Config.save()` writes the complete configuration to `~/.nex-voice/config.json` as plaintext. 4. The file receives permissions based on the process `umask` or retains unsafe pre-existing permissions. 5. An attacker with local filesystem access reads the configuration file. 6. The attacker extracts the API key and uses it against the configured LLM provider. This issue requires local access or another process executing under an account that can read the file; it does not independently provide remote acc ...[truncated 602 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
lib/action_extractor.py:288
Finding

Arbitrary LLM Endpoint Configuration Can Disclose API Credentials and Full Transcripts

Content
View full analysis
None: """Set API base URL""" self.set("api_base", base) def get_api_model(self) -> str: """Get API model""" return self.get("api_model", AI_MODEL) ``` ```python response = subprocess.run( [ "curl", "-s", f"{api_base}/chat/completions", "-H", f"Authorization: Bearer {api_key}", "-H", "Content-Type: application/json", "-d", json.dumps({ "model": model, "messages": [ { "role": "system", "content": "You are an expert at extracting action items from transcripts. Always return valid JSON.", }, { "role": "user", "content": prompt, } ], "temperature": 0.7, }) ], capture_output=True, text=True, ) ``` The prompt embedded in this request includes the complete transcript: ```python prompt = f"""Extract action items from this transcript. For each action, identify: - type: task, reminder, call, email, meeting, decision, or deadline - description: what needs to be done - assigned_to: who is responsible (if mentioned) - due_date: when it's due (ISO format YYYY-MM-DD or full timestamp, or null) - priority: low, medium, or high Return a JSON array of action items. Transcript: {text} JSON output:""" ``` ### Technical Analysis The application permits an arbitrary string to be stored as `api_base`. When LLM extraction is requested, `curl` sends both of the following to that destination: - The configured API key in an `Authorization: Bearer` header. - The complete recording ...[truncated 2417 chars]
Remediation
View remediation

T08 · Insecure Dependencies

Note
Location
README.md:16
Finding

Unpinned Third-Party Dependency Installation Instructions

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
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
Findings (36)

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

The supplied code is a narrow transcription utility, not the broader productivity system described. The transcription portion does align with the declared purpose: it uses Whisper, supports configurable language/model, and keeps processing local via CLI tools. However, the major claimed capabilities—intelligent action extraction, searchable transcript storage, and action-item management—are absent from this code chunk. While audio conversion via ffmpeg is an extra implementation detail rather than a harmful undeclared capability, the overall description materially overstates what this code actually does, so this is a mismatch.

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 30)May include surrounding context.

pip install faster-whisper

text
3. Install FFmpeg:
- Ubuntu/Debian: `sudo apt-get install ffmpeg`
- macOS: `brew install ffmpeg`
- Windows: `scoop install ffmpeg` or `choco install ffmpeg`

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The README advertises LLM-based action extraction as optional but does not explicitly state that enabling --use-llm transmits transcript content, which may contain sensitive meeting notes, personal reminders, or business data, to a third-party provider. In a voice transcription skill whose primary privacy claim is that data stays on the user's machine, this omission can mislead users into exposing sensitive content without informed consent.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
92% confidence
Finding

The skill advertises and instructs use of shell execution, local file reads/writes, environment-backed configuration, and setup scripts, but it does not declare any explicit tool scope or allowed-tools boundary. In an agent environment, this weakens least-privilege controls and can allow broader-than-expected command and filesystem access when the skill is invoked.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The trigger phrases are broad, generic, and overlap with normal conversation such as 'transcribe', 'action item', or 'what did we discuss', increasing the chance of accidental invocation. In an agentic setting, unintended activation can cause processing of sensitive local audio/transcript data or shell-backed actions without clear user intent.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
86% confidence
Finding

The skill supports optional configuration of external LLM providers and explicitly references a remote API endpoint, which creates a real data-transmission path for transcripts and derived notes. Because transcripts may contain sensitive personal or business information, sending them to third-party APIs can create confidentiality, compliance, and retention risks if users are not clearly warned and consent is not explicit at execution time.

Content

Scanner excerpt · SKILL.md (reported line 176)May include surrounding context.

nex-voice config set-api-key nex-voice config set-provider openai nex-voice config set-model gpt-4o nex-voice config set-api-base https://api.openai.com/v1

text

## Example Interactions

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

When use_llm is enabled, the function sends the full transcript to an external API endpoint, which may include sensitive voice-note or meeting content. This file contains no consent, warning, minimization, or destination restriction logic, so confidential information can be disclosed to third parties or attacker-controlled endpoints via api_base.

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 · lib/action_extractor.py (reported line 288)May include surrounding context.

python
JSON output:"""

    try:
        response = subprocess.run(
            [
                "curl",
                "-s",

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The skill's purpose is transcription and action extraction, and optional LLM integration is declared in the manifest, so contacting an LLM API is in scope. However, invoking an external process to do this adds a shell/tool execution capability that is not justified by the stated purpose of a voice-note organizer and is materially broader than ordinary text processing.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The configuration defaults WHISPER_LANGUAGE to "nl", which imposes a specific language setting unless the user overrides it. This is a natural-language locale preference embedded in code without any visible opt-in or justification that the skill is region-specific.

Content

No source excerpt is available for this finding.

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 · README.md (reported line 142)May include surrounding context.

md
# LLM settings (optional)
AI_API_KEY = os.getenv("AI_API_KEY")
AI_API_BASE = os.getenv("AI_API_BASE", "https://api.openai.com/v1")
AI_MODEL = os.getenv("AI_MODEL", "gpt-4o")

# Audio limits

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 · lib/config.py (reported line 88)May include surrounding context.

python
# LLM settings (optional)
AI_API_KEY = os.getenv("AI_API_KEY")
AI_API_BASE = os.getenv("AI_API_BASE", "https://api.openai.com/v1")
AI_MODEL = os.getenv("AI_MODEL", "gpt-4o")

# Audio limits

Tainted flow: 'CONFIG_FILE' from os.getenv (line 21, credential/environment) → open (file write)

Medium
Category
Data Flow
Confidence
65% confidence
Finding

Data from a source is assigned to a variable that is later passed to a sink, creating a variable-mediated taint flow.

Content

Scanner excerpt · lib/config.py (reported line 114)May include surrounding context.

python
def save(self) -> None:
        """Save configuration to file"""
        with open(CONFIG_FILE, 'w') as f:
            json.dump(self.config_data, f, indent=2)

    def get(self, key: str, default: Any = None) -> Any:

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

set_api_key stores the API key in config_data and save() writes that data to config.json in plaintext. On multi-user systems, backups, synced home directories, or malware-assisted local access, this can expose the credential and enable unauthorized use of the user's LLM account.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The schema sets the default recording language to 'nl', which imposes a specific locale by default. The file does not show any user choice, opt-in, or documented regional constraint that would justify forcing Dutch as the default language.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The function signature defaults the language argument to 'nl', which creates a language policy constraint in code. There is no indication here that users can opt into this locale or that the skill is limited to a Dutch-only context.

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 · lib/transcriber.py (reported line 16)May include surrounding context.

python
def check_whisper() -> bool:
    """Check if whisper CLI is available"""
    try:
        subprocess.run(
            [WHISPER_CMD, "--version"],
            capture_output=True,
            check=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 · lib/transcriber.py (reported line 29)May include surrounding context.

python
def check_ffmpeg() -> bool:
    """Check if ffmpeg is available"""
    try:
        subprocess.run(
            ["ffmpeg", "-version"],
            capture_output=True,
            check=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 · lib/transcriber.py (reported line 42)May include surrounding context.

python
def get_audio_duration(file_path: str) -> float:
    """Get audio duration in seconds using ffmpeg"""
    try:
        result = subprocess.run(
            [
                "ffmpeg",
                "-i", file_path,

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/transcriber.py (reported line 76)May include surrounding context.

python
output_path = input_path.parent / f"{input_path.stem}_converted.{output_format}"

    try:
        subprocess.run(
            [
                "ffmpeg",
                "-i", str(input_path),

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

Audio conversion always uses '-y' and writes to a deterministic sibling filename, so an existing file can be silently overwritten. In a voice-note application handling user recordings and transcripts, this can cause unintended data loss or replacement of prior converted outputs without clear user awareness.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
83% confidence
Finding

The function defaults to a configured WHISPER_LANGUAGE and always passes '--language' to the transcription command when no explicit user language is supplied. This can impose a fixed language/locale behavior without offering user choice or documenting a justified region-specific constraint.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

Transcription writes a JSON sidecar containing transcript data and segments to disk in the same directory as the source audio before deleting it. For sensitive voice notes, even temporary plaintext transcript artifacts can be exposed to other local users, backup agents, sync tools, or crashes that prevent cleanup, increasing confidentiality risk despite the product's privacy-focused context.

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 · lib/transcriber.py (reported line 134)May include surrounding context.

python
]

    try:
        result = subprocess.run(
            cmd,
            capture_output=True,
            text=True,

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The transcribe command silently duplicates the user-provided audio file into AUDIO_DIR, creating persistent local storage beyond the original input file. While the code later prints transcript results, there is no prior disclosure or confirmation that the original audio will be retained in a separate application-managed location.

Content

No source excerpt is available for this finding.