The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructions mandate reading the nearest CONTEXT.md file to preserve project terms and explain the frontier. This creates a surface for indirect prompt injection if the CONTEXT.md file is sourced from an untrusted third party or contains malicious instructions intended to subvert the agent's behavior.
Ingestion points: The skill reads CONTEXT.md in SKILL.md (section 'Explain The Frontier').
Boundary markers: No explicit delimiters or warnings to ignore instructions within CONTEXT.md are provided.
Capability inventory: The skill itself does not perform dangerous operations like file writes or network requests, but acts as a control flow mechanism.
Sanitization: No sanitization of the content of CONTEXT.md is mentioned.

what