adaptation

Warn

Audited by Gen Agent Trust Hub on Mar 11, 2026

Risk Level: MEDIUMPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The skill describes an indirect prompt injection surface where the agent's system prompt is rewritten based on external history or feedback.
  • Ingestion points: Untrusted 'task_history' and performance reports are processed by an 'optimizer' in the adaptation_loop function within SKILL.md.
  • Boundary markers: The implementation lacks delimiters or 'ignore embedded instructions' warnings for the feedback history.
  • Capability inventory: The save_prompt() function allows the system to persist and apply modified instructions to its core configuration.
  • Sanitization: No sanitization or validation of the proposed prompt is shown before persistence.
  • [COMMAND_EXECUTION]: The skill proposes a 'Code Evolution' use case where an agent 'refactors' its own helper functions at runtime. This pattern of self-modifying code creates a critical risk if the refactoring logic can be influenced by malicious input strings within the task history. Evidence found in SKILL.md under the 'Use Cases' section.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Mar 11, 2026, 07:00 AM