recursive-training

Pass

Audited by Gen Agent Trust Hub on May 13, 2026

Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill instructs the agent to execute several local scripts (e.g., recursive-training-grpo.py, recursive-training-loader.py) located in the .recursive/scripts/ folder. These scripts are responsible for orchestrating memory extraction and retrieval across runs.\n- [PROMPT_INJECTION]: The skill ingests and processes all markdown files within the /.recursive/run/ directory to generate memory items. This creates a surface for indirect prompt injection where malicious content in previous run artifacts could be extracted and persisted into the agent's long-term memory shards.\n
  • Ingestion points: Any markdown artifact within completed run folders, including plan summaries and memory impact reports.\n
  • Boundary markers: No specific delimiters or safety instructions are defined in the architectural references to separate training data from operational instructions.\n
  • Capability inventory: Local script execution (Python/PowerShell) and the ability to update memory shards within the repository structure.\n
  • Sanitization: The documentation does not specify any sanitization, validation, or filtering of the content extracted from the processed markdown files.
Audit Metadata
Risk Level
SAFE
Analyzed
May 13, 2026, 04:58 PM
Security Audit — agent-trust-hub — recursive-training