The Agent Skills Directory

[DYNAMIC_EXECUTION]: The MCPAgent class implementation includes a create_tool method that generates Python code via an LLM, writes it to the local filesystem, and then loads it for execution by connecting to it as a new server. This creates a significant risk of arbitrary code execution if the tool description is derived from untrusted input.
[COMMAND_EXECUTION]: The skill defines patterns for executing system commands, including a run_command tool and a SandboxedExecution class. The sandboxing implementation uses subprocess.run with shell=True, which is susceptible to command injection if command strings are not meticulously sanitized.
[EXTERNAL_DOWNLOADS]: The ContextManager class provides a pattern for fetching content from arbitrary external URLs using the requests.get method to augment the agent's context.
[INDIRECT_PROMPT_INJECTION]: The skill documents an architecture that ingests untrusted data from external URLs and files while maintaining access to high-privilege tools (file writing, command execution), creating a vulnerability surface for indirect prompt injection.
Ingestion points: ContextManager.add_url and ContextManager.add_file methods in SKILL.md.
Boundary markers: None are present in the provided design patterns to distinguish between instructions and data.
Capability inventory: run_command, write_file, edit_file, and dynamic tool creation/execution.
Sanitization: While basic path validation is included in SandboxedExecution, no sanitization or escaping is applied to content fetched from URLs or generated by the LLM.

autonomous-agent-patterns