output-regression-doctor

Pass

Audited by Gen Agent Trust Hub on Oct 18, 2026

Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The test suite utilizes shell command execution to validate the CLI entry point.
  • Evidence: tests/test_image_diff.py contains subprocess.run calls to execute the compare_image.py script using the local Python interpreter. This is standard testing behavior and not present in the core skill logic.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted image data (metadata/text) which could potentially contain malicious instructions.
  • Ingestion points: scripts/image_diff.py reads image files.
  • Boundary markers: SKILL.md contains explicit instructions to ignore embedded instructions in images.
  • Capability inventory: The skill can write files to the local system.
  • Sanitization: scripts/report.py implements a safe() function to escape Markdown special characters in the output report, preventing rendering-based injection attacks.
  • [SAFE]: The skill implements significant defensive measures against resource exhaustion and denial-of-service attacks.
  • Evidence: scripts/image_diff.py enforces a 50 MiB file size limit and a 16 million pixel limit, and it explicitly enables Pillow's decompression bomb warnings as errors.
  • [SAFE]: The skill prevents unauthorized file modification through safe file-handling practices.
  • Evidence: scripts/compare_image.py uses the xb (exclusive binary write) and x (exclusive text write) flags, ensuring the tool will not overwrite existing files or previous reports.
  • [SAFE]: No network access or data exfiltration vectors were identified.
  • Evidence: The script imports and logic are limited to local file system and image processing operations; no networking libraries are imported.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 18, 2026, 01:19 PM