The Agent Skills Directory

[DATA_EXFILTRATION]: The script scripts/detect_mcp_servers.py reads the user's global configuration file ~/.claude.json. This file typically contains sensitive information, including API tokens and environment variables for Model Context Protocol (MCP) servers. While the script filters the output to only return server names, reading this file exposes credentials to the execution context.
[DYNAMIC_EXECUTION]: The utility scripts/write_bruhs_config.py uses subprocess.run to dynamically invoke sync_bruhs_block.py using the current Python executable. This pattern of dynamic script invocation increases the complexity of the execution environment and could be exploited if inputs were not properly sanitized.
[EXTERNAL_DOWNLOADS]: The commands/walk.md workflow includes instructions to download HTML, CSS, and JavaScript assets from a remote GitHub repository (github.com/cursor/plugins) at runtime using curl. These assets are used to render an interactive PR walkthrough, representing a remote code execution vector.
[INDIRECT_PROMPT_INJECTION]: The skill has a significant attack surface for indirect prompt injection. Commands like /bruhs:peep and /bruhs:cook ingest untrusted data from external sources.
Ingestion points: Pull request comments and bodies are fetched via gh api in commands/peep.md; Linear ticket descriptions are fetched via MCP in commands/cook.md.
Boundary markers: The prompts used for sub-agents (e.g., in peep.md) use Markdown headers like ## PR Context and ## Review Comment to separate data from instructions.
Capability inventory: Sub-agents used to process this data have access to powerful tools including Bash, Edit, Read, and Task.
Sanitization: The skill implements a verification loop where agents must reproduce bugs with tests before reporting them, which serves as a functional validation step.

bruhs