persona-memory

Pass

Audited by Gen Agent Trust Hub on Mar 10, 2026

Risk Level: SAFEPROMPT_INJECTIONEXTERNAL_DOWNLOADSDATA_EXFILTRATION
Full Analysis
  • [PROMPT_INJECTION]: The skill implements an indirect prompt injection surface where content from local memory blocks (~/.ai-memory/blocks/*.md) is retrieved and appended to the agent's system prompt at runtime. • Ingestion points: Untrusted data enters the agent context through the remember.ts script. • Boundary markers: The launch-claude.sh script uses triple-dash delimiters (---) to separate the persona profile from the memory context. • Capability inventory: The skill can execute the agent binary and manage local memory files. • Sanitization: No sanitization of the retrieved memory strings is performed before they are interpolated into the system prompt.
  • [EXTERNAL_DOWNLOADS]: The evaluation dashboard references font assets from trusted Google domains (fonts.googleapis.com, fonts.gstatic.com).
  • [DATA_EXFILTRATION]: The serve-eval-dashboard.ts script accesses local shell configuration files (~/.zprofile, ~/.env) to retrieve the ANTHROPIC_API_KEY for use with the official Anthropic API during testing. This sensitive file access is part of the intended developer convenience for the skill's evaluation dashboard.
Audit Metadata
Risk Level
SAFE
Analyzed
Mar 10, 2026, 03:57 AM