iterative-retrieval
Pass
Audited by Gen Agent Trust Hub on Mar 20, 2026
Risk Level: SAFEDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
- [DATA_EXFILTRATION]: Reference to sensitive local configuration directory.
- Evidence: The skill explicitly lists 'Agent definitions in ~/.claude/agents/' in its 'Related' section in SKILL.md.
- Impact: This directory typically contains internal system prompts and tool configurations for the agent. If an agent following this retrieval pattern attempts to ingest these files as 'context', it could lead to the exposure of internal configuration data and system instructions.
- [PROMPT_INJECTION]: Indirect prompt injection surface identified in the context retrieval pattern.
- Ingestion points: The skill describes a loop for retrieving content from codebase files (e.g., 'src/**/*.ts') and evaluating it for relevance in SKILL.md.
- Boundary markers: No delimiters or explicit instructions to ignore embedded prompts are defined for the retrieved file content.
- Capability inventory: The pattern assumes the agent possesses file-reading and search capabilities.
- Sanitization: The methodology does not include validation or sanitization of content retrieved from the codebase before it is presented to the agent for evaluation.
Audit Metadata