The Agent Skills Directory

[DYNAMIC_EXECUTION]: The ReAct pattern example includes a calculator tool that uses the Python eval() function on user-supplied strings. This allows for the execution of arbitrary Python code if the input expression is not strictly validated or sandboxed.
[DYNAMIC_EXECUTION]: The documentation provides a run_notebook helper function that utilizes nbconvert.preprocessors.ExecutePreprocessor to programmatically execute all code cells within a Jupyter notebook, which constitutes full remote code execution if the target notebook is modified or sourced from an untrusted location.
[COMMAND_EXECUTION]: The DualMemoryAgent implementation constructs Neo4j Cypher queries using Python F-strings (e.g., in add_semantic_fact and retrieve_semantic). This pattern is vulnerable to Cypher injection if the entity names or relationship types are derived from untrusted external data.
[INDIRECT_PROMPT_INJECTION]: Several architectures, such as the research_agent and Meta-Controller, ingest data from external search tools (Tavily) and interpolate the raw results directly into LLM prompts. This creates an attack surface where malicious instructions embedded in web content could potentially override agent behavior.
Ingestion points: research_agent node in SKILL.md (uses TavilySearchResults), research_specialist in SKILL.md (uses TavilySearchResults).
Boundary markers: Absent; external content is appended to prompts using basic string interpolation.
Capability inventory: The skill includes tools for eval() execution, file system access via notebooks, and network searches.
Sanitization: No explicit sanitization or instruction-ignoring delimiters are used for external search results.
[EXTERNAL_DOWNLOADS]: The skill instructs the user to clone a repository from GitHub (github.com/FareedKhan-dev/all-agentic-architectures.git) and install multiple third-party dependencies from public registries.

all-agentic-architectures