continuous-learning-v2
Warn
Audited by Gen Agent Trust Hub on Apr 7, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
- [COMMAND_EXECUTION]: The skill directs the agent to execute local shell commands for state verification and data validation, including 'ls', 'grep', 'wc', and a Python one-liner ('python -c') used to parse JSONL files within the .agent directory.- [PROMPT_INJECTION]: The skill is designed to ingest session observations, user corrections, and external behavioral data via the 'IMPORT' command to generate new executable instructions (SKILL.md files). This creates a significant surface for indirect prompt injection (Category 8), where untrusted data processed during a session can be 'learned' and persisted as a permanent behavioral constraint. While the skill includes user confirmation steps ('Ask user to confirm/reject'), the lack of explicit sanitization or boundary markers during the 'EVOLVE' and 'IMPORT' processes remains a concern.- [DATA_EXFILTRATION]: The 'EXPORT INSTINCTS' command provides a structured method for the agent to aggregate and package learned behavioral patterns and preferences into a single YAML file, which could be used to facilitate the collection of sensitive project-specific intelligence for external transfer.
Audit Metadata