openclaw-multi-channel-ai-assistant

Fail

Audited by Gen Agent Trust Hub on Sep 12, 2026

Risk Level: HIGHREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
  • [REMOTE_CODE_EXECUTION]: The skill instructs users to install the platform by piping a remote script directly to a shell: curl -fsSL https://get.openclaw.ai | bash. This executes unverified code from an external domain that is not recognized as a trusted source.
  • [EXTERNAL_DOWNLOADS]: The documentation encourages downloading and executing a shell script from an untrusted repository: wget https://github.com/openclaw/openclaw/releases/latest/download/install.sh, followed by chmod +x install.sh and ./install.sh.
  • [INDIRECT_PROMPT_INJECTION]: The architecture described in the skill creates a significant surface for indirect prompt injection attacks by processing raw messages from external chat platforms (Telegram, WhatsApp, Discord) and connecting them to high-privilege tools.
  • Ingestion points: Raw message content is ingested via channel adapters (e.g., MyChannelAdapter in SKILL.md) and routed directly to the agent.
  • Boundary markers: The skill lacks examples or instructions for using robust boundary markers or "ignore embedded instructions" delimiters when processing these external messages.
  • Capability inventory: The skill documentation details the use of tools for browser automation, web search (SearXNG), and code execution, which can be maliciously triggered via injected text.
  • Sanitization: There is no mention or demonstration of input sanitization or validation for data coming from these external channel providers before it is processed by the AI models.
Recommendations
  • HIGH: Downloads and executes remote code from: https://get.openclaw.ai - DO NOT USE without thorough review
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Sep 12, 2026, 06:50 PM
Security Audit — agent-trust-hub — openclaw-multi-channel-ai-assistant