self-improving-agent-remember

Pass

Audited by Gen Agent Trust Hub on Jul 18, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The skill implements a persistent memory system (MEMORY.md) that stores user-provided information for future sessions, creating a surface for indirect prompt injection.\n
  • Ingestion points: User input from the /si:remember command is processed and written to a local memory file.\n
  • Boundary markers: The skill lacks instructions or delimiters for the agent to treat the memory file contents as untrusted data or to ignore embedded commands.\n
  • Capability inventory: The skill uses shell commands (grep, sed, pwd) and file-write operations; the broader agent context involves reading this file at startup.\n
  • Sanitization: No sanitization or escaping of the user-provided content is performed before storage, although the instructions warn against storing credentials.\n- [COMMAND_EXECUTION]: The duplicate check workflow uses a shell command (grep -ni "") that interpolates user-derived keywords. If the agent does not properly sanitize these keywords, it could lead to arbitrary command execution via shell metacharacters.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 18, 2026, 04:19 PM
Security Audit — agent-trust-hub — self-improving-agent-remember