The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides implementation patterns for autonomous agents that ingest and act upon untrusted external data, creating a potential surface for indirect prompt injection.
Ingestion points: Untrusted data enters the agent context through variables such as query in the ReAct pattern example and goal or objective in the Plan-Execute logic within references/detailed-guide.md.
Boundary markers: The provided code scaffolds do not implement explicit delimiters or boundary markers (e.g., XML tags or clear 'ignore instructions' prefixes) to isolate untrusted input from the agent's core instructions.
Capability inventory: The guide describes agents with capabilities for tool execution (executor.execute), external API interactions (RobustAPIClient), and potentially sensitive file operations as outlined in the TASK_PERMISSIONS dictionary.
Sanitization: While the documentation conceptually recommends 'Input validation' and 'Action constraints' in the 'Guardrailed Autonomy' section, the provided code examples do not include active sanitization or filtering of the ingested content.

autonomous-agents