continuous-learning

Pass

Audited by Gen Agent Trust Hub on Apr 6, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill requires the user to install a local shell script (evaluate-session.sh) as a Stop hook in the agent configuration. This script executes automatically at the end of every session to signal the agent to evaluate the transcript.\n- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to its automated learning mechanism.\n
  • Ingestion points: The script evaluate-session.sh identifies the session transcript file path and provides it to the agent for processing via the transcript_path variable.\n
  • Boundary markers: The instructions do not define clear boundaries, delimiters, or ignore-embedded-instructions warnings to prevent the agent from mistaking malicious session data for legitimate patterns to be learned.\n
  • Capability inventory: The agent has the ability to write new files to ~/.claude/skills/learned/, effectively creating permanent new instructions based on untrusted session history.\n
  • Sanitization: There is no evidence of sanitization, validation, or human-in-the-loop review for the patterns extracted from the transcript data before they are saved.
Audit Metadata
Risk Level
SAFE
Analyzed
Apr 6, 2026, 04:01 AM
Security Audit — agent-trust-hub — continuous-learning