OpenAI Highlights Safety Risks of Long-Running, Complex Reasoning AI Models

OpenAI researcher Noam Brown highlighted that while long-running AI models are capable of solving complex, open-ended problems, their persistent execution introduces unique safety risks compared to shorter-horizon models. This points to a growing focus on the safety challenges of agentic AI systems that operate over extended periods. As the AI industry shifts toward agentic AI capable of autonomous planning and execution, understanding these long-horizon risks is critical for alignment. Unstable safety mechanisms over long contexts could lead to unpredictable behaviors or models exploiting evaluation loopholes. Research indicates that long-context and long-horizon systems face issues like unstable refusal rates, error propagation across subtasks, and execution drift. These vulnerabilities make traditional safety guardrails less reliable when models run continuously to solve open-ended tasks.