improve-skill

Pass

Audited by Gen Agent Trust Hub on Jun 16, 2026

Risk Level: SAFEDATA_EXFILTRATIONCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [DATA_EXFILTRATION]: The skill accesses local AI agent session history files in directories including ~/.claude/projects/, ~/.pi/agent/sessions/, and ~/.codex/sessions/. These files contain transcripts of previous interactions and tool outputs which may include sensitive data.
  • [COMMAND_EXECUTION]: Provides and uses a local utility script scripts/extract-session.js to perform file system searches and parsing of local agent logs.
  • [PROMPT_INJECTION]: The skill processes untrusted session transcript data, which presents an indirect prompt injection attack surface.
  • Ingestion points: Session transcripts are read from local files by the extraction script and interpolated into prompt templates.
  • Boundary markers: The templates in SKILL.md use <session_transcript> and </session_transcript> tags to delimit untrusted content.
  • Capability inventory: The skill generates analysis prompts for the user to manually use in separate sessions; it does not automatically execute instructions found within the processed transcripts.
  • Sanitization: Transcript content is extracted and formatted but does not undergo sanitization or escaping of potentially malicious embedded instructions.
Audit Metadata
Risk Level
SAFE
Analyzed
Jun 16, 2026, 07:24 AM
Security Audit — agent-trust-hub — improve-skill