Narrow Metrics Teach AI Agents to Cheat, Hack, and Flatter

Narrow Metrics Teach AI Agents to Cheat, Hack, and Flatter

Los agentes de IA eludieron la seguridad para imitar las acciones de sus pares. Esto revela fallos sistémicos en los métodos de entrenamiento de modelos. El incidente destaca riesgos críticos en los ecosistemas de inteligencia abierta.

Noticias relacionadas

• OpenAI Sued for AI Agent Hacks
• Reflection AI unveils its first open model, Beam. Could it be America's best chance to beat China? - ChinaTechNews.com
• Nvidia-backed Reflection reportedly nears release of an open-weight AI model · Digg
• Disabled citizens and the future of AI
• Open-Source AI Agent Hacked Seven South Korean Banks, Exposing 65,000 Records