LATIDIA · Ciberseguridad

Configuración, no conciencia: un estudio empírico a gran escala de las indicaciones del sistema LLM

arXiv: 2609.31575v1Tipo de anuncio: nuevo Resumen: Las indicaciones filtradas del sistema a menudo se tratan como ventanas a los valores ocultos de los modelos de lenguaje comercial, sin embargo, su composición rara vez se estudia a escala. Analizamos una fusión

La noticia

arXiv:2609.31575v1 Announce Type: new Abstract: Leaked system prompts are often treated as windows into the hidden values of commercial language models, yet their composition is rarely studied at scale. We analyze a merged corpus of 407 leaked, reconstructed, or officially published system prompts from 62 vendors across four community collections, identifying 29 near-duplicate clusters covering 66 files. Operational content rather than ethical statements dominates the corpus; a deliberately simple block-level classifier assigns roughly 58\% of classified words to tool/protocol and roughly 5\% to safety policy, while the strictest rule-lines guard tool use and file safety over harmful content by an 11:1 margin. Literal text transfer concentrates in a small set