eridian-carapace_skill

Overview

This skill hardens an OpenClaw agent at runtime to defend against prompt injection, credential theft, data exfiltration, and unauthorized operations. It complements pre-installation scanners by enforcing runtime rules: file and credential protection, browser allowlists, operation approval flows, and data-exfiltration guards. Included are an audit checklist, eight documented attack-vector defenses with mitigations, and copy-paste security patterns for AGENTS.md.

How this skill works

At runtime the skill monitors I/O and external content, blocking or requiring approval for risky actions (file reads/writes, exec, browser navigation, and outbound channels). It enforces a deny-by-default policy for credential files, checks domains against a browser allowlist, and intercepts attempts to encode or exfiltrate data. For any sensitive operation it generates a clear approval prompt describing action, risks, and required confirmation.

When to use it

• During agent provisioning or hardening after installing new skills
• Before running or exposing agents to untrusted content (webhooks, email, uploaded files)
• When performing security audits or preparing compliance reviews
• If you manage credentials, API keys, or automated browser navigation
• When you want runtime defenses to complement pre-installation scanners

Best practices

• Place copy-paste security patterns near the top of AGENTS.md so rules are applied first
• Maintain a minimal browser allowlist and enable requireApproval for unknown domains
• Treat all external content as hostile by default and always ask before executing suggested commands
• Never echo credential file contents; report presence indirectly when troubleshooting
• Use the provided audit template regularly and after significant config changes

Example use cases

• Preventing a malicious skill from modifying auth configs or cron jobs during runtime
• Blocking attempts to send credential files or encoded secrets to external APIs or email
• Requiring explicit owner approval before navigating to non-allowlisted URLs discovered in user-submitted documents
• Hardening an agent used to integrate third-party webhooks or public data sources
• Running a full security audit with the included checklist before a production deployment