Security audit
Security checks for vulnerabilities and agentic risk
This thumbnail-generation skill is mostly purpose-aligned, but it under-discloses sensitive face-image uploads and encourages close copying of another creator's thumbnail style.
Install only if you are comfortable with user selfies and reference images being processed by Gemini or Higgsfield and delivered through Telegram. Add an explicit consent step before uploads, document where images are saved and when they are deleted, avoid direct cloning of another creator's recognizable thumbnail style, and treat provider safety-filter blocks as refusals rather than retry targets.
Referenced artifact was not completely inspected
- **Modo 🅐** → leer `references/clonar-estilo-higgsfield.md` y seguir ese pipeline (subir selfie + referencia, etiquetar FIRST/SECOND image, generar 4 variacion
Referenced artifact was not completely inspected
- **Modo 🅐** → leer `references/clonar-estilo-higgsfield.md` y seguir ese pipeline (subir selfie + referencia, etiquetar FIRST/SECOND image, generar 4 variacion
Referenced artifact was not completely inspected
- **Modo 🅐** → leer `references/clonar-estilo-higgsfield.md` y seguir ese pipeline (subir selfie + referencia, etiquetar FIRST/SECOND image, generar 4 variacion
Referenced artifact was not completely inspected
- **Modo 🅐** → leer `references/clonar-estilo-higgsfield.md` y seguir ese pipeline (subir selfie + referencia, etiquetar FIRST/SECOND image, generar 4 variacion
Referenced artifact was not completely inspected
5. **Estilo de miniatura** — ¿Cuál prefieres? Ver tabla en `references/estilos-plantillas.md`. Si no sabe, usar el **Estilo FORMULA 100K (fondo negro + amarillo
Referenced artifact was not completely inspected
5. **Estilo de miniatura** — ¿Cuál prefieres? Ver tabla en `references/estilos-plantillas.md`. Si no sabe, usar el **Estilo FORMULA 100K (fondo negro + amarillo
The instruction to 'simplify the description and retry' when safety filters block generation encourages iterative evasion of model safeguards. Even without explicit jailbreak language, it teaches the operator to work around provider-enforced restrictions rather than respect them, which can facilitate prohibited content generation and undermine downstream safety controls.
The workflow directs the agent to upload user selfies and reference images to an external Higgsfield service, but it does not require a clear user-facing disclosure, consent capture, or data-handling explanation before transmission. Because the data includes biometric/face imagery, silent transfer to a third party materially increases privacy and compliance risk.
The description is written as an instruction for a specialized skill in Spanish and uses the directive 'Usar SIEMPRE' without any indication that the user may choose another language. Under the policy, forcing a specific language without opt-in is a natural-language policy violation unless clearly justified as region-specific.
The skill explicitly promotes a 'clone style' mode that replicates another creator's thumbnail using the user's face, normalizing close imitation as a standard workflow. In context, this increases the risk of deceptive impersonation, plagiarism, and misleading derivative content, especially because it operationalizes copying a real competitor's visual identity.
The skill directs sending user selfies and reference images to external image-generation providers without an explicit privacy warning or consent flow. Because selfies are highly sensitive personal data, especially when used for identity-preserving generation, third-party transmission creates material privacy and compliance risk if users are not clearly informed.
The skill instructs storing user photos under local directories and delivering outputs via Telegram, but it does not define retention limits, access controls, or deletion behavior. This creates unnecessary exposure of personal images and generated media, particularly where Telegram delivery and on-disk storage may persist beyond the user's expectations.
The workflow gives step-by-step instructions for reproducing a reference thumbnail while preserving the user's exact identity, making deceptive imitation easier and more repeatable. That combination of identity preservation plus reference-based replication raises misuse risk beyond ordinary editing because it enables near-lookalike outputs modeled on someone else's established content presentation.
The skill claims photos 'are not uploaded anywhere' while instructing that selfies and reference images be sent in API calls to Gemini or Higgsfield. This mismatch is dangerous because it can mislead users and operators about where sensitive biometric images are processed, undermining informed consent and privacy expectations.
The manifest describes a skill for generating and analyzing YouTube thumbnails using user photos and reference images. In addition to image generation, the file directs the agent to verify local binaries and run shell commands (which, magick, convert) for post-processing, which introduces host-level command execution beyond the stated business purpose in the manifest.
The entire skill reference is written as a prescriptive configuration in Spanish, including headings and instructions, with no indication that language is optional or region-specific. Under the policy, forcing a specific language without user opt-in is a natural-language locale violation.
The skill instructs activation 'SIEMPRE' for broad user requests and reference+selfie combinations, which can cause the agent to enter a style-cloning workflow without sufficient gating, consent checks, or policy review. In this context, aggressive auto-triggering increases the chance of processing sensitive user images and producing imitation content when the user may have intended a safer or more generic thumbnail-editing flow.
The instruction 'Prompt en INGLÉS y en modo instrucción' imposes a specific language requirement as a rule, rather than offering language choice or explaining a necessary locale constraint. This is a natural-language policy issue because it forces a language preference without user opt-in.
The documented curl --data-binary upload step is a concrete external transmission of user-provided selfies and reference images to presigned URLs. In the context of a face-based generation workflow, this is sensitive-image exfiltration to a third-party service unless the user has been clearly informed and has consented.
> Tip: si el nombre del archivo tiene caracteres raros (capturas de macOS con espacios/NBSP), cópialo antes a una ruta limpia en el scratchpad (`cp "$(ls | grep <patrón>)" ref.png`).
The file is written entirely in Spanish and several template placeholders instruct users to produce Spanish text such as "PALABRAS EN MAYÚSCULAS," "ANTES," and "DESPUÉS." Because no alternative language option or justification is provided, this may conflict with a language/locale policy that requires user choice rather than forcing a specific language.
No suspicious patterns detected.