The Agent Skills Directory

[PROMPT_INJECTION]: No malicious prompt injection patterns or attempts to override agent safety guidelines were detected. The instructions include explicit reminders that the host platform's permissions and authorization remain authoritative.
[DATA_EXFILTRATION]: No evidence of unauthorized data access or exfiltration. Network operations (via curl) are used exclusively for fetching generated media assets from tool-provided URLs (Higgsfield, Monid) as part of the intended workflow. No sensitive file paths or environment variables are accessed.
[REMOTE_CODE_EXECUTION]: The skill relies on local CLI tools (ffmpeg, higgsfield, monid, codex) already present on the host. It does not perform dangerous operations like downloading and piping remote scripts to a shell (e.g., curl | bash).
[OBFUSCATION]: The skill contains no obfuscated code, Base64-encoded commands, zero-width characters, or homoglyph-based attacks. All scripts and instructions are in plain text.
[INDIRECT_PROMPT_INJECTION]: While the skill ingests user input for content generation, it uses structured prompt templates and defines clear boundaries for the AI's role, minimizing the surface for indirect injection attacks.
[DYNAMIC_EXECUTION]: No dangerous dynamic code execution patterns (like eval or exec on untrusted input) were found. The Python and JavaScript components use standard libraries and APIs for image processing and DOM manipulation.

scroll-world