Security audit
Security checks for vulnerabilities and agentic risk
This skill is a disclosed media-workflow helper that uses a third-party image generator and local output files, with some scope and environment caveats but no evidence of deception or unsafe persistence.
Install only if you are comfortable sending transcript-derived prompts to Higgsfield and saving generated PNGs locally. For OpenClaw use, provide a timed transcript as requested, confirm credit cost before generation, and treat the local ffmpeg/uvx steps as non-OpenClaw reference workflow.
The trigger text uses an overly broad activation rule (o cualquier variación) that can cause the skill to auto-run on loosely related requests. In an agentic environment, broad invocation criteria increase the chance of the wrong skill handling user content, which can lead to unintended external generation, file creation, or processing costs without clear user intent.
Lines L013-L015 explicitly state that steps 1-3 do not run in this OpenClaw environment and that job_display and open "$DEST" do not apply. However, later sections L042-L073 and L183-L185 still document running those local commands and UI actions as if they were part of the skill flow, creating a direct contradiction in the skill's operational instructions.
The phrase acompáñame la toma is ambiguous and may match benign requests that are not asking for this specific workflow. Ambiguous triggers are dangerous in automation because they can route conversations into media-processing and paid-image-generation steps the user did not clearly authorize.
uvx/uv tool run commands without ==version create a rug-pull risk.
La instrucción establece de forma obligatoria un único registro lingüístico para todo contenido generado, sin ofrecer elección al usuario ni documentar una justificación de cumplimiento regional. Esto constituye una restricción de idioma/locale impuesta por defecto.
The skill instructs downloading externally hosted image URLs with curl and saving outputs locally, but it does not visibly warn the user that data will be exchanged with a third-party service and retrieved from external URLs. In this context the danger is unauthorized or unclear third-party data flow: user-derived content is sent to Higgsfield and externally hosted results are fetched, which can create privacy, compliance, and consent issues.
The manifest description frames the skill as starting from a provided transcript with timestamps and generating scrapbook graphics, emphasizing that it does not edit the video. Yet the main pipeline in L042-L073 includes validating a video file, extracting audio with ffmpeg, and transcribing it with mlx-whisper, which materially broadens the behavior beyond operating on an already-supplied transcription.
No suspicious patterns detected.
