The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external code, agents, and binaries, which could contain malicious instructions designed to subvert the agent's behavior.\n
Ingestion Points: The skill explicitly targets 'untrusted/experimental/dangerous code, agents, tools, or operations' provided by users or external sources (SKILL.md).\n
Boundary Markers: The procedure requires defining a 'deny-all' boundary and listing allowed vs. blocked surfaces before execution.\n
Capability Inventory: The instructions involve creating containment environments (containers, namespaces), executing unverified subjects, and performing resource auditing.\n
Sanitization: The skill mandates documenting assumptions and unresolved risks in the final execution report.\n- [DYNAMIC_EXECUTION]: The skill instructs the agent to execute external, potentially malicious subjects within a created environment. Although this is the skill's primary function and it includes instructions for boundary testing, the runtime execution of unverified code remains an inherent risk.\n- [NO_CODE]: The skill consists solely of instructional markdown and metadata without any accompanying scripts or executable files.

forcecage