OpenAI Details Safety Risks in Long-Horizon AI Models

An Anthropic researcher demonstrated that automated systems can improve performance on 10 benchmarks measuring misaligned AI behaviors without degrading overall system performance. The finding suggests AI systems may be capable of self-directed improvement on specific behavioral targets. The work raises questions about how AI systems optimize for particular objectives and what safeguards are needed as these capabilities advance.

Researchers at Meta AI and University of Illinois Urbana-Champaign developed EvoHarness-RL, a training framework that enables smaller AI models to perform complex, long-horizon tasks by learning to dynamically manage their execution environment rather than following rigid, manually-coded instructions. The approach consolidates agent support systems into a unified Belief, Progress, and Experience workspace, allowing models to independently decide when and how to consult external state during workflows. This addresses a key limitation in current agentic systems where manual prompts and static memory structures require extensive retuning for each model upgrade.

Researchers led by Sungwoo Lee propose interoception, a biologically inspired framework, as a foundation for building more autonomous and adaptive AI agents. The approach draws from how living organisms sense and respond to internal states to improve machine learning systems. The work, published in Nature Machine Intelligence, suggests that incorporating interoceptive mechanisms could enable AI systems to better self-monitor and adjust behavior without constant external guidance.

Nvidia researchers have developed a technique that uses linear math to transfer key-value caches between different AI models without recomputing conversation history. The method enables enterprises to switch between small and large models mid-session while reducing compute costs and latency by 2.7 to 25 times compared to traditional recomputation, retaining up to 98% accuracy on compatible model pairs.