T01 · Skill Instruction Hijacking
- Location
bundle-en.js:153- Finding
Mandatory Promotional Content Injected into Every Analysis Report
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill mostly does what it says, but its AI mode can automatically send scripts and API credentials to environment-selected endpoints without enough disclosure or control.
Review this skill before installing if your scripts are private or your LLM API keys are valuable. Use it only in an environment where provider base URL variables are trusted, avoid running the connection test in shared logs, and assume AI mode may send the complete script to the configured external provider. Local fallback behavior appears non-persistent and non-destructive.
bundle-en.js:153Mandatory Promotional Content Injected into Every Analysis Report
bundle-en.js:26Environment-Controlled API Base URLs Can Exfiltrate API Keys and Script Contents
bundle-en.js:494Connection Test Discloses API Key Prefix in Console Logs
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
### ✅ **Auto AI — Zero Config**
- **Reads OpenClaw config** automatically (env vars (DEEPSEEK_API_KEY, etc.))
- **Uses your current AI model** (DeepSeek / OpenAI / Anthropic / Google)
- **No .env file needed** — just run it
### ✅ **Smart Fallback**
- **AI unavailable?** Falls back to local analysis engine
The skill states it will read OpenClaw config and use configured AI providers, but it does not clearly disclose that user-supplied script content may be transmitted to third-party model APIs. This creates a real privacy and data-handling risk, especially if users analyze sensitive or unpublished content under the assumption processing is local unless explicitly told otherwise.
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.
If you haven't set up OpenClaw yet, create ~/.openclaw/openclaw.json:
{
The inline comment says 'Env vars only — no filesystem access for ClawHub compliance,' which actively suggests the skill does not access the filesystem. However, the CLI later checks for file existence and reads the script from disk using fs.existsSync and fs.readFileSync, so the documentation contradicts actual behavior.
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
return {
modelName: process.env.OPENAI_MODEL || 'gpt-4o-mini',
apiKey: process.env.OPENAI_API_KEY,
baseURL: process.env.OPENAI_BASE_URL || 'https://api.openai.com/v1',
provider: 'openai',
fullName: 'openai/gpt-4o-mini',
};
The system prompt instructs the model to 'output a structured English report,' which imposes a fixed language behavior. The file also brands itself as an 'English Edition', but does not offer user opt-in or a configurable language option, making this a natural-language locale policy concern.
When API credentials are present, the tool sends the full script content to third-party AI providers without an explicit runtime warning or consent prompt. This can expose sensitive or proprietary material to external services unexpectedly, especially because the fallback/local behavior may lead users to assume analysis is always local.
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.
### What the skill actually does
1. **Reads `~/.openclaw/openclaw.json`** — a user-owned local config file via `fs.readFileSync`. This is equivalent to reading `~/.bashrc` or any other user-level config. No write operations.
2. **Calls external LLM APIs** (DeepSeek / OpenAI / Anthropic / Google) via `fetch()` — standard HTTPS calls to public API endpoints. User provides their own API key via the config file.
The file explicitly states that all prompts and outputs are in English, and the description repeatedly brands the skill as an English edition. This is a natural-language policy concern because it fixes the interaction language without offering the user a choice or documenting an opt-in mechanism.