other
Warning
- Location
- SKILL.md:140
- Finding
- Fabricated First-Hand Product Testing and Unsupported Evidence Claims## Vulnerability Details **File Location**: `SKILL.md:140`; corroborating templates at `references/ejemplos-reviews.md:56,127` and `references/estructura-detallada.md:33` **Vulnerability Type**: Deceptive Content Generation **Risk Level**: Medium ### Complete Vulnerable Snippets `SKILL.md:140`: ```markdown - **Experiencia**: Escribe como si hubieras probado el producto. Usa frases como "tras probarlo durante 3 semanas" o "en nuestra prueba de rendimiento" ``` `references/ejemplos-reviews.md:56`: ```markdown Si el Dyson se escapa de presupuesto, el Rowenta X-Force Flex 14.60 es la alternativa más sólida del mercado por la mitad del precio. Ofrece una succión excelente para el uso cotidiano y su brazo flexible es uno de los mejores que hemos probado para limpiar debajo de los muebles sin agacharse. ``` `references/ejemplos-reviews.md:127`: ```markdown - "Tras X semanas de uso, podemos confirmar que…" ``` `references/estructura-detallada.md:33`: ```markdown > Elegir una aspiradora sin cable puede ser frustrante. Hay decenas de modelos en el mercado, las fichas técnicas son confusas y las reviews en Amazon no siempre son fiables. Para ahorrarte horas de investigación, hemos analizado más de 20 modelos y seleccionado los 8 mejores según potencia de succión, autonomía real y facilidad de uso. Tanto si buscas el modelo más completo como el más económico, aquí encontrarás tu respuesta. ``` ### Technical Analysis The skill explicitly directs the agent to write as though it had personally tested a product, including claims about a three-week trial or an internal performance test. Supporting templates reinforce this behavior with statements that products were tested or that more than twenty models were analyzed. These instructions can cause simulated or nonexistent experience to be represented as genuine first-hand evidence. This conflicts with the skill's separate direction to use real data and avoid invented c ...[truncated 1849 chars]
- Remediation
- ## Remediation Suggestions 1. Replace the instruction to “write as if” a product was tested with an explicit prohibition against claiming unperformed testing or personal use. 2. Permit first-person experience statements only when the user supplies genuine, attributable test notes or other verifiable evidence. 3. Require research-based assessments to be labeled clearly, such as “Based on manufacturer specifications and cited customer feedback.” 4. Require citations or `[VERIFY]` placeholders for prices, availability, specifications, warranties, ratings, and review counts. 5. Remove unsupported first-person claims and concrete metrics from reusable examples, or mark them prominently as fictional placeholders. 6. Add a pre-publication verification checklist covering factual claims, source dates, affiliate disclosures, and evidence for personal-experience statements. 7. Instruct the agent to distinguish manufacturer claims, third-party test results, customer opinions, and the publisher's own verified observations. 8. Fail safely when evidence is unavailable: use neutral wording and disclose that the product was not personally tested rather than inferring experience.