pixelmator-pro
Pass
Audited by Gen Agent Trust Hub on Oct 1, 2026
Risk Level: SAFECOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill uses
subprocess.runacross multiple Python scripts (pxm-batch.py,pxm-compose.py,pxm-info.py) to execute local system utilities includingbash,osascript, and the macOS image processing toolsips. These commands are used to implement the core functionality of the skill and do not target external or untrusted binaries. - [DYNAMIC_EXECUTION]: In
pxm-compose.py, the skill generates AppleScript code at runtime based on a JSON specification provided to the script. It also utilizes a bash wrapper (pxm.sh) to execute AppleScript and JXA (JavaScript for Automation) from stdin or local files. The implementation includes a dedicated sanitization functions(text)to escape backslashes and quotes in user-provided strings, mitigating AppleScript injection risks. - [INDIRECT_PROMPT_INJECTION]: The skill is designed to process external image files and JSON specifications, which constitutes a surface for indirect prompt injection if the data source is untrusted.
- Ingestion points:
pxm-batch.py(processes image files and directories),pxm-compose.py(processes JSON layout specifications). - Boundary markers: The skill uses structured JSON for its internal logic but does not explicitly instruct the AI agent to ignore potentially malicious instructions embedded in the data it processes (e.g., in image metadata or text layers).
- Capability inventory: The skill possesses the capability to execute shell commands, run AppleScript, write to the local file system (for image exports), and read system-level image metadata.
- Sanitization: String inputs for dynamic code generation are escaped, and numeric or hex-based inputs (like colors) are validated against strict patterns using regular expressions.
- [PRIVILEGE_ESCALATION]: The skill requires macOS Automation and Accessibility permissions to interact with the Pixelmator Pro app and its system dialogs. These permissions are managed by the operating system's TCC (Transparency, Consent, and Control) framework and require explicit user approval via standard OS prompts during the first execution.
Audit Metadata