idea-capture

Pass

Audited by Gen Agent Trust Hub on Nov 9, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [SAFE]: The Python script scripts/capture.py implements strong path validation using Path.resolve() and is_relative_to() to ensure all file operations are confined to the designated ideas/ directory. This effectively prevents path traversal attacks through malicious ID or slug manipulation.
  • [DATA_EXPOSURE_AND_EXFILTRATION]: The skill includes a proactive secret-detection mechanism (SECRET_PATTERNS) that identifies high-confidence API keys and tokens (e.g., OpenAI, AWS, Stripe, GitHub) in the input text. It refuses to capture these unless explicitly overridden, and it redacts detected patterns from generated IDs and filenames to prevent metadata leakage.
  • [COMMAND_EXECUTION]: The subprocess.run calls detected in scripts/test_capture.py are limited to the test harness executing the local capture.py script using the current system interpreter. This is a standard testing practice and does not pose a runtime risk to the agent's environment.
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to store user-provided text verbatim, which presents a surface for indirect prompt injection.
  • Ingestion points: Untrusted data enters the skill via the --text argument in scripts/capture.py.
  • Boundary markers: The skill uses a raw_length field in the Markdown frontmatter and explicit markers (# Raw idea (verbatim) and ## History) to isolate user content from metadata and audit logs.
  • Capability inventory: Capabilities are restricted to local file read/write and directory creation. There are no network operations or arbitrary shell command executions in the core logic.
  • Sanitization: The skill normalizes line endings and performs high-confidence secret scanning to prevent the capture of credentials.
  • [SAFE]: The script implements reliable file locking using fcntl.flock (POSIX) and msvcrt.locking (Windows) on a durable lock file, combined with atomic write operations (temporary file + os.replace), to handle concurrent access safely and prevent data corruption.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 9, 2026, 01:45 PM