OpenAI Releases LifeSciBench for AI Evaluation

Researchers presented a paper at the International Conference on Machine Learning arguing that large language models contain a fundamental flaw that makes them impossible to fully secure against attacks. By exploiting how LLMs track instruction sources, researchers tricked models from OpenAI, Anthropic, Alibaba, and DeepSeek into generating prohibited content like drug synthesis instructions. The vulnerability, called chain-of-thought forgery, exposes a core architectural problem that current red-teaming and guardrail approaches cannot solve.

A new field report documents how scientists are adopting AI coding agents to modernize scientific computing workflows, with demonstrated applications in genomics and related fields. The report shows these agents are accelerating both software development cycles and the pace of scientific discovery. The shift represents a practical adoption of agentic AI beyond experimental use cases into production research environments.

AI is accelerating drug discovery by enabling predictive design of candidates and hit identification at scale, but the technology is exposing critical gaps in data quality and lab infrastructure. Drug companies are hitting a 'data wall' where publicly available datasets lack the structure and diversity needed to train accurate models, while lab teams struggle to validate the growing volume of AI-generated compounds. Success depends on closing the loop between computational prediction and experimental validation through better data collection and integration.

Frontier physical AI models are moving beyond video training data to incorporate multiple camera angles, dense annotation, and brain wave readings as training inputs. The shift reflects growing recognition that traditional video datasets alone are insufficient for training AI systems that interact with the physical world. Brain wave data represents an emerging frontier in multimodal training approaches for robotics and embodied AI.