The Agent Skills Directory

[COMMAND_EXECUTION]: The skill implements a multi-harness-orchestrator which is designed to emit and execute shell commands. It specifically mentions generating commands such as claude -p "$(cat CLAUDE.md; cat SHARING.md) <task>" to delegate work to other agent environments.
[INDIRECT_PROMPT_INJECTION]: The skill is configured to automatically ingest and prioritize instructions from project-level files such as GROK.md, CLAUDE.md, and AGENTS.md. This creates a vulnerability surface where malicious instructions placed in these files by an external contributor or attacker could influence the agent's behavior.
Ingestion points: Processes content from CLAUDE.md, SHARING.md, AGENTS.md, and GROK.md located within the active project directory.
Boundary markers: The skill body does not specify the use of delimiters or 'ignore' instructions to separate repository data from core system prompts.
Capability inventory: The skill possesses the ability to execute shell commands, perform directory and file operations, and interact with GitHub via MCP tools.
Sanitization: No sanitization or validation routines are described for the data read from these configuration files before it is interpolated into prompts for other models.
[DATA_EXFILTRATION]: The orchestration logic involves reading the entire contents of local files (like SHARING.md and CLAUDE.md) and passing that data as a prompt to an external AI service. While intended for context building, this pattern facilitates the transmission of project-specific metadata to external providers.

grok-harness