The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill includes several Python scripts (agent_evaluator.py, agent_planner.py, tool_schema_generator.py) that ingest data from external JSON files. This creates a potential vector where untrusted data (such as execution logs or requirement descriptions) could influence the agent's behavior during analysis or design phases.
Ingestion points: agent_evaluator.py (Line 485), agent_planner.py (Line 448), and tool_schema_generator.py (Line 434) read and process data from user-supplied JSON files using json.load().
Boundary markers: The scripts lack explicit boundary markers or instructions to the LLM to ignore potentially adversarial directives embedded within natural language fields like task_description or error_message.
Capability inventory: The toolkit allows agents to write generated architectures, diagrams, and evaluation reports to the local filesystem using configurable output paths.
Sanitization: While the scripts perform basic structural validation of JSON data, they do not sanitize natural language content that is subsequently processed by the agent.

agent-designer