LATIDIA · Ciberseguridad

Ajuste fino de autorreflexión: mejora de la seguridad del agente contra ataques de inyección rápida por experiencia de falla

arXiv:2610.04269v1 Announce Type: cross Abstract: Los agentes del modelo de lenguaje grande (LLM) se implementan cada vez más en entornos aumentados por herramientas, pero su dependencia de entradas externas los hace altamente vulnerables a prompt i

La noticia

arXiv:2610.04269v1 Announce Type: cross Abstract: Large language model (LLM) agents are increasingly deployed in tool-augmented environments, but their reliance on external inputs makes them highly vulnerable to prompt injection attacks that can hijack task objectives. Existing safety alignment methods rely on static expert trajectories or preference optimization, limiting their ability to generalize to adaptive attack patterns. In this work, we propose Self-Reflection Fine-Tuning (SRFT), a training framework that enables agents to improve robustness by learning from their own failure experiences under adversarial conditions. Instead of passively imitating expert behaviors, SRFT exposes the agent to compromised trajectories constructed via injected attacks, and leverages an expert model to generate structured self-reflection reasoning