The Agent Skills Directory

[DYNAMIC_EXECUTION]: Multiple language templates provided in the skill implement a "calculator" tool using unsafe dynamic execution methods. In templates/genaiscript/autonomous_agent.genai.mts, the tool uses eval(args.expression), and in templates/typescript/parallelization.ts, it uses the Function constructor as an evaluator. These methods are susceptible to arbitrary code execution if the input strings are not strictly validated, as they run with the privileges of the agent's environment.
[COMMAND_EXECUTION]: The Python autonomous agent template (templates/python/autonomous_agent.py) includes a run_command tool that utilizes subprocess.run(command, shell=True). While the implementation includes a command whitelist (ls, cat, grep, etc.), the use of shell=True combined with user-supplied arguments is a well-known security risk for shell injection, especially if the agent is not strictly sandboxed.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to build agents that ingest untrusted data from various sources (RAG, web results, tool outputs), creating a significant attack surface for indirect prompt injection.
Ingestion points: resources/augmented-llm.md (retrieval systems), resources/dynamic-orchestration.md (autonomous exploration/web search results), and various worker templates.
Boundary markers: The skill explicitly instructs users to use XML-style delimiter tags (e.g., <retrieved_context>) to separate untrusted data from system instructions.
Capability inventory: The provided templates include powerful capabilities such as file system operations (read_file, write_file), command execution, and network access via the Anthropic API.
Sanitization: While the documentation provides guidance on sanitizing role-switching patterns and jailbreak phrases before injection, the provided templates do not implement these filters directly, relying instead on user implementation of the described architectural patterns.

agent-patterns