Security audit
Security checks for vulnerabilities and agentic risk
The skill mostly matches its stated video-generation purpose, but it deserves review because it can inspect and upload personal reference photos and persist photo-linked media IDs for future reuse without enough per-use control.
Install only if users understand that personal reference photos may be uploaded to Higgsfield and their media IDs may be reused later. Prefer providing one specific approved photo, confirm costs before generation, avoid giving access to broad selfie folders, and ask the agent not to save or publish media IDs unless reuse is explicitly desired.
The activation trigger includes "any variation" combined with broad long-script/video phrasing, which can cause the skill to fire on loosely related requests. Over-broad invocation increases the risk of unintended tool use, unnecessary media processing, and access to user content when a more limited or different skill should have handled the request.
The 'Cuándo activar' section uses generic phrases like requests for visual resources and cinematic clips, which may overlap with ordinary editing, ideation, or asset-planning tasks. In context, that can lead the agent to invoke costly external generation workflows or request extra user assets without sufficiently clear user intent.
Las instrucciones de preguntas al usuario están definidas únicamente en español y no ofrecen alternativa de idioma. Según la política, forzar un idioma o locale sin elección explícita del usuario constituye una violación de política lingüística.
The skill instructs the agent to inspect a local folder of personal selfies and select among multiple private photos, which expands data access beyond what is strictly necessary to fulfill a B-roll request. Even if the goal is legitimate avatar generation, broad access to personal image repositories increases the chance of over-collection and unintended exposure of sensitive biometric or personal images.
The skill explicitly directs the agent to store and reuse a user's reference-photo media identifier across future sessions, creating persistent linkage to biometric-like personal media. Cross-session retention without clear consent, retention limits, or revocation controls increases privacy risk and the chance of unauthorized reuse.
Including the exact reference-photo media ID in the final deliverable unnecessarily exposes an identifier tied to the user's uploaded reference image. That identifier could be reused, shared, or logged in downstream systems, increasing the risk of unintended access or correlation of personal media across projects.
The instruction actively claims the agent can 'leer 3-4 archivos con Read y elegir' the best photo based on visual characteristics like face angle and lighting. That contradicts the actual capability implied by the tool name and typical behavior, because evaluating image composition is not something plain file reading accomplishes.
La instrucción prescribe una caracterización demográfica específica en lenguaje natural para todos los clips con avatar. Aunque puede reflejar un caso de uso concreto, aquí aparece como mandato general de la skill sin justificación contextual ni opción de ajuste, lo que puede entrar en conflicto con políticas organizacionales sobre atributos sensibles en instrucciones predefinidas.
No suspicious patterns detected.
