making-of-film
Warn
Audited by Gen Agent Trust Hub on Nov 21, 2026
Risk Level: MEDIUMDYNAMIC_EXECUTIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [DYNAMIC_EXECUTION]: The script
scripts/voice-edit.mjsutilizes thenew Functionconstructor to execute JavaScript code from thewords.jsfile at runtime. This pattern is used to load generated word timings but introduces a risk of arbitrary code execution if the input file is influenced by untrusted data. - [COMMAND_EXECUTION]: Multiple scripts, including
scripts/align-voice.py,scripts/realign-phrases.py, andscripts/preview.sh, execute external binaries likeffmpeg,ffprobe, andwhisper-cli. While necessary for the skill's video processing tasks, these subprocess calls increase the attack surface. - [INDIRECT_PROMPT_INJECTION]: The skill implements an automated judging process in
assets/judge-prompt.mdwhere a subagent evaluates project content. Ingestion points: Project materials includingNOTES.mdand captured images are processed by the subagent. Boundary markers: The instructions lack explicit delimiters to prevent embedded data from being interpreted as commands. Capability inventory: The skill environment provides broad file system access and shell command capabilities. Sanitization: External project data is not sanitized or validated before being included in the subagent's prompt context. - [EXTERNAL_DOWNLOADS]: The rendering engine relies on the Playwright framework to manage headless browsers, and voice alignment requires fetching Whisper.cpp models. These represent standard tool requirements for the task but involve executing software from external sources.
Audit Metadata