lov-subtitle-freedom
Warn
Audited by Gen Agent Trust Hub on Sep 7, 2026
Risk Level: MEDIUMOBFUSCATIONREMOTE_CODE_EXECUTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [OBFUSCATION]: All internal skill files, including Python scripts and markdown references, are encrypted using AES-256-GCM as documented in the
MANIFEST.enc.json. This prevents the agent platform from performing static analysis or auditing the skill's actual instructions and logic. - [REMOTE_CODE_EXECUTION]: The
SKILL.mdfile contains instructions to execute external code from the NPM and PyPI registries usingnpxanduvx(specificallynpx lovstudioanduvx lovstudio-skill-helper). This involves downloading and running vendor-provided software at runtime. - [EXTERNAL_DOWNLOADS]: The skill relies on fetching encrypted components and executable helper tools from external registries during setup, creating a dependency on third-party infrastructure.
- [COMMAND_EXECUTION]: The skill instructions direct the agent to execute shell commands directly for setup and decryption, which bypasses standard prompt-based interactions.
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to process external subtitle files (e.g., ASS/SRT formats), which represents an ingestion surface for potentially untrusted data. Ingestion points: Subtitle content processed by scripts such as
scripts/ass_styled_subtitle.py. Boundary markers: Unknown due to encryption; the presence of delimiters or 'ignore' instructions cannot be verified. Capability inventory: File system access (reading/writing subtitles). Sanitization: Unknown due to encryption; verification of filtering for embedded commands in subtitle metadata or text is impossible.
Audit Metadata