improve-skill

Warn

Audited by Gen Agent Trust Hub on Mar 8, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill executes a local script scripts/extract-session.js to extract data from agent session history.- [DATA_EXFILTRATION]: The skill accesses sensitive directory paths including ~/.claude/projects/ and ~/.codex/sessions/ to read session transcripts. These transcripts frequently contain proprietary code, environment details, and potentially sensitive credentials or data discussed during agent sessions.- [PROMPT_INJECTION]: The skill processes untrusted data from session history, creating an indirect prompt injection surface.
  • Ingestion points: Transcripts extracted from local JSONL logs are directly interpolated into new prompts.
  • Boundary markers: The prompt template uses simple <session_transcript> tags which can be bypassed by malicious content in the session history.
  • Capability inventory: The agent is subsequently instructed to write files to the filesystem based on the analysis of the untrusted transcript.
  • Sanitization: No validation or sanitization is performed on the extracted content to prevent embedded instructions from being executed by the agent in the new session.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Mar 8, 2026, 08:01 AM