The Agent Skills Directory

[REMOTE_CODE_EXECUTION] (CRITICAL): Piped remote script execution found in references/install/nodejs.md. The instruction curl -sL [URL] | bash downloads and executes a shell script directly from a non-trusted GitHub repository (Syncause/ts-agent-file), allowing for arbitrary code execution.
[CREDENTIALS_UNSAFE] (HIGH): Hardcoded GitHub Personal Access Tokens (split into syncause.repo.token.p1 and syncause.repo.token.p2) are embedded in references/install/java.md. These credentials provide unauthorized access to the author's private repository assets.
[EXTERNAL_DOWNLOADS] (HIGH): The skill downloads and installs binary artifacts and configuration files from non-trusted external sources, including Python wheels and Node.js packages (@syncause/debug-mcp), without integrity checks or version pinning.
[COMMAND_EXECUTION] (MEDIUM): The skill identifies application entry points and automatically injects initialization code. It also dynamically generates and executes reproduction scripts (reproduce_issue.py) via subprocess calls based on user-provided symptoms.
[DATA_EXFILTRATION] (LOW): Instrumentation captures runtime method snapshots (arguments, local variables, and logs) and transmits them to an external endpoint (wss://api.syn-cause.com). This exposure is consistent with the skill's purpose but poses a risk to sensitive data.
[PROMPT_INJECTION] (LOW): Mandatory Evidence Chain for Indirect Prompt Injection: (1) Ingestion point: SKILL.md uses user symptoms to generate logic. (2) Boundary markers: Absent. (3) Capability inventory: Subprocess calls in Phase 1 and file writes across all install guides. (4) Sanitization: Absent. The skill is vulnerable to instructions embedded in issue descriptions.

syncause-debugger