photocraft

Pass

Audited by Gen Agent Trust Hub on Oct 23, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDYNAMIC_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes complex, attacker-controllable data formats such as PSD, PSB, and JSON action sets, which presents a surface for indirect prompt injection. Maliciously crafted files could attempt to influence the agent's logic or bypass intended constraints.
  • Ingestion points: Data enters the agent's context through file opening commands (app.open, photocraft-cli convert), image processing tools, and the ingestion of batch action files (batch, droplet).
  • Boundary markers: The skill explicitly instructs the agent to use --automation-read-root and --automation-write-root flags to establish a directory-level security boundary for file operations.
  • Capability inventory: The skill allows for engine command execution, file system writes within the specified roots, and interaction with the system print spooler via lp or Windows-specific commands.
  • Sanitization: A comprehensive "fail-closed" security architecture is implemented, as verified in crates/automation/src/workspace.rs, which explicitly blocks ambient filesystem access and restricts sensitive command parameters.
  • [COMMAND_EXECUTION]: The skill relies on the execution of shell-based commands for its core functionality, including application management and CLI interactions.
  • Evidence: Instructions include the use of open -a PhotoCraft, photocraft-cli subcommands, and system-level printing via lp or the Windows spooler.
  • [EXTERNAL_DOWNLOADS]: The skill references downloading binary components and source code from an external repository not included in the pre-defined trusted organizations list.
  • Evidence: The skill directs users to the GitHub releases page for the storytold organization (github.com/storytold/photocraft) for downloading the CLI and other assets.
  • [DYNAMIC_EXECUTION]: The skill supports the installation and execution of WebAssembly-based plugins, which constitutes runtime dynamic code execution.
  • Evidence: PhotoCraft implements a Wasm plugin ABI (v1) using the wasmi interpreter. The implementation includes significant resource constraints, such as a 50M-instruction fuel limit and no access to network or filesystem APIs, to mitigate the risks associated with untrusted modules.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 23, 2026, 02:52 PM