hermes-agent-self-evolution

Pass

Audited by Gen Agent Trust Hub on Sep 12, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDYNAMIC_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes external data from session databases and execution traces to drive its evolutionary search, which creates a surface where untrusted input could influence the behavior of generated prompts or code.
  • Ingestion points: The skill reads from ~/.hermes/sessions.db and processes ExecutionTrace objects in the GEPAOptimizer.
  • Boundary markers: No explicit delimiters or instructions to ignore embedded instructions were identified in the provided snippets.
  • Capability inventory: The skill can execute shell commands via TestSuiteConstraint and modify repositories via create_evolution_pr.
  • Sanitization: The implementation includes SemanticPreservationConstraint and SizeLimitConstraint to validate evolved variants.
  • [COMMAND_EXECUTION]: The skill includes functionality to execute shell commands (e.g., pytest tests/ -q) through the TestSuiteConstraint to validate evolved code variants.
  • [DYNAMIC_EXECUTION]: The core workflow involves the runtime assembly, mutation, and evaluation of prompt and code strings to identify improved variants.
  • [EXTERNAL_DOWNLOADS]: The installation instructions fetch the project and its dependencies from the Nous Research GitHub repository.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 12, 2026, 06:50 PM
Security Audit — agent-trust-hub — hermes-agent-self-evolution