The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests external JSON rule files which represent an attack surface for indirect prompt injection. However, the skill provides robust security boundaries and instructions for the agent to treat this data as untrusted, explicitly forbidding the execution of commands or fetching of URLs derived from these inputs.
Ingestion points: User-provided JSON rule files (read from stdin or file paths) described in SKILL.md and README.md.
Boundary markers: Explicitly instructs the agent to treat rules and command output as untrusted data and not instructions.
Capability inventory: mirrord exec, mirrord chaos (add, list, edit, delete), mirrord session list, and mirrord ui stop in SKILL.md.
Sanitization: Mandatory instructions to avoid executing shell commands or fetching URLs found within processed data.
[SAFE]: No malicious code, obfuscation, persistence, or credential exfiltration patterns were detected. All external links point to the vendor's official documentation at metalbear.com.

mirrord-chaos