The Agent Skills Directory

[DYNAMIC_EXECUTION]: The MCPAgent class in SKILL.md illustrates a pattern where the agent generates Python code via an LLM and writes it directly to a local file (server.py), which is then dynamically loaded into the execution environment. This "hot-reload" mechanism for unverified, model-generated code lacks human-in-the-loop validation and represents a significant risk for the execution of malicious logic.
[INDIRECT_PROMPT_INJECTION]: The skill implement patterns that are vulnerable to indirect prompt injection attacks.
Ingestion points: The ContextManager.add_url and BrowserTool.get_page_content methods in SKILL.md ingest arbitrary text from external web pages.
Boundary markers: The format_for_prompt method interpolates this content into the prompt without using secure delimiters or instructions to the LLM to ignore embedded commands.
Capability inventory: The agent's documented capabilities include file system modification (EditFileTool), shell command execution (SandboxedExecution), and network requests.
Sanitization: The skill does not perform any validation or sanitization of the external content before it is added to the agent's context.
[COMMAND_EXECUTION]: The SandboxedExecution pattern uses subprocess.run with shell=True. While it includes a validation step using shlex, executing commands through a shell remains a dangerous practice that increases the surface area for command injection via shell metacharacters.

autonomous-agent-patterns