evolving-ai-agents
Pass
Audited by Gen Agent Trust Hub on Oct 29, 2026
Risk Level: SAFECOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The framework's default evolution engine, AEvolveEngine (SkillForge), is designed to use an LLM with bash tool access to perform file-system operations. This allows the system to programmatically read, edit, and manage workspace files including system prompts and skill documents as part of its optimization loop.
- [DYNAMIC_EXECUTION]: A-Evolve supports the runtime creation and modification of executable content. The engine can mutate Python tool implementations (e.g., in the tools/ directory) and reload agent state, including prompts and skills, from the filesystem during execution to apply evolved logic.
- [INDIRECT_PROMPT_INJECTION]: The skill describes an autonomous feedback loop where data from external benchmark tasks is ingested by an agent and subsequently analyzed by an evolution LLM. This architecture presents an inherent surface for indirect prompt injection, as adversarial content within benchmark tasks could theoretically influence the evolution engine's decisions when it modifies the agent's core instructions.
Audit Metadata