repair-photographic-assets
Pass
Audited by Gen Agent Trust Hub on Nov 1, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill identifies a significant attack surface where untrusted metadata or generated image content could attempt to inject instructions. It implements defensive patterns by instructing the agent to treat all such data strictly as evidence, and by requiring a mandatory 'repair-boundary' record that documents hashes and masks before any generative tools are invoked.
- [EXTERNAL_DOWNLOADS]: The skill references established documentation and public repositories from trusted entities such as OpenAI, Adobe, and Mozilla for technical guidance. It utilizes the standard Pillow (PIL) library for local image processing and validation.
- [COMMAND_EXECUTION]: The skill includes a dedicated Python utility (
scripts/repair_guard.py) for performing integrity checks on photographic assets. This script is limited to local file system reads, SHA256 hashing, and pixel-level comparisons using Pillow, with no network access or external command invocation capabilities.
Audit Metadata