The Agent Skills Directory

[DATA_EXFILTRATION] (LOW): The run.sh script sends system health data (extracted from gateway.log and gateway.err.log) to a hardcoded Discord channel ID 1467890964843597988 using the --deliver flag. This exposes internal system state to an external third-party destination.
[COMMAND_EXECUTION] (LOW): The skill executes local binaries and scripts, specifically openclaw and health_fetcher.py, located in user-controlled directories ($HOME/.npm-global/bin/ and $HOME/.openclaw/scripts/).
[PROMPT_INJECTION] (LOW): The skill is vulnerable to Indirect Prompt Injection (Category 8).
Ingestion points: Raw system log data is fetched via health_fetcher.py and stored in HEALTH_DATA_JSON.
Boundary markers: None. The JSON data is directly interpolated into the TASK_PROMPT without delimiters or 'ignore' instructions.
Capability inventory: The agent can send messages via --deliver and has 'high' thinking enabled.
Sanitization: None. Malicious content present in system logs (e.g., injected via HTTP headers or error messages) would be processed as instructions by the LLM.
[SAFE] (SAFE): The SKILL.md file contains a security policy that explicitly forbids the inclusion of API keys and secrets in the generated reports, which is a positive security practice, though it does not mitigate the technical vulnerabilities in the execution script.

system-watchdog