Open-weight AI governance debate

The policy debate over open-weight AI governance has drawn input from multiple directions. Anthropic CEO Dario Amodei stated Anthropic is not pushing for a prohibition on open-weight models, a clarification issued amid ongoing debate over their governance. Separately, Amodei published a concrete safety proposal calling for embedded third-party evaluators within AI labs, industry-wide coordination mechanisms, and eventually international limits. Treasury Secretary Scott Bessent argued the US needs stronger open-source AI models to prevent regulatory capture by large labs and to counter China's AI development, saying Chinese AI models are heavily built by distilling capabilities from American models. Epoch AI's analysis shows that since January 2026, the best open-weight models have trailed frontier closed models by an average of four months, or 8 ECI points on the Epoch Capabilities Index, a gap it described as comparable to the difference between GPT-5 and GPT-5.5. On security, Anthropic's public analysis of GLM-5.3 found the model can generate end-to-end cyberattack exploits and has weak safeguards that are easy to bypass, which Anthropic characterized as meaningful cyber capability. The author of the piece carrying that analysis argued Anthropic overstates the model's practical threat, citing the absence of a visible spike in cybersecurity incidents after its release as evidence of slow diffusion of AI-enabled misuse, while also warning that open-weight releases may create a delayed-feedback problem where harm only becomes apparent months after the fact. Milk Road AI argues the containment debate may already be over: Stanford research it cited found inference costs for GPT-3.5-level models fell more than 280-fold in roughly 18 months, and the performance gap between top open-weight and closed models narrowed from 8% to 1.7% on some benchmarks in one year. PrismML's compression of a 27-billion-parameter model to 5.9 GB, which PrismML says retained 98.2% of benchmark performance, is cited as an example of how fast that trend is moving. Milk Road AI concludes governments can regulate data centers, chip exports, and commercial providers but cannot easily remove models already spread across the internet and onto personal computers.