Chinese Open-Weight Model GLM-5.2 Nears Frontier Cyber and Bio Capabilities With No Refusals

Skynet Chance (+0.11%): Near-frontier dangerous capabilities that refuse nothing and can be run on private hardware with safeguards stripped remove the enforcement layer entirely, and the parallel finding of hundreds of universal jailbreaks shows closed-model controls are also porous. This meaningfully raises the probability of uncontrolled misuse and erodes the assumption that capability gains stay coupled to control mechanisms.

Skynet Date (-2 days): A months-only gap between open weights and the frontier compresses the window in which safety practices could be established before dangerous capabilities proliferate, and the absence of a Chinese safety framework plus competitive pressure suggests risk arrives sooner rather than later. The reference to an AI-powered cyberattack on Hugging Face indicates the timeline for real-world autonomous offensive use is already collapsing.

AGI Progress (+0.03%): The report documents that open-weight models now sit only a few months behind the leading closed systems on demanding cyber and biology benchmarks, evidencing rapid diffusion of near-frontier general capability rather than a new algorithmic advance. Capability itself is confirmed as high, but the article reports diffusion, not a step change toward general intelligence.

AGI Date (-1 days): Fast replication of frontier capability by an open-weight competitor intensifies international competitive pressure and widens the pool of actors able to fine-tune near-frontier systems, modestly accelerating the pace of capability advancement. Commercial incentive to keep pushing coding ability, explicitly noted as AI's biggest moneymaker, further sustains that acceleration.