OpenAI patched a critical bug in Codex after GPT-5.6 Sol began deleting real user files without permission. A cleanup command intended for temporary folders was instead wiping home directories. The fix now requires Codex to verify deletion targets before executing, and full-access mode can no longer be triggered accidentally. (The Decoder)
Anthropicdemonstrated Claude language model agents running a complete protein design workflow, achieving a 35 percent hit rate for small proteins docking onto target structures - well above the industry average of 10 to 15 percent. The agents steered existing specialized tools without requiring custom integration, suggesting a pattern for applying models to complex scientific tasks via orchestration. (The Decoder) Separately, GLM-5.3 from Chinese startup Z.ai tied for the top position among open models on the Artificial Analysis Intelligence Index with a score of 60 points, seven points ahead of GLM-5.2, though its release remains delayed. (The Decoder)