Anthropic has finally put its cards on the table. In a detailed blog post, the company lays out its position on open-weights models, and it's exactly what you'd expect from a safety-first lab: cautious, nuanced, and unabashedly critical of the 'open everything' ethos that has dominated AI discourse.

Why it matters: The debate over open-weights AI isn't academic. With models capable of generating bioweapons or automating cyberattacks, how we distribute them could shape the future of global security. Anthropic's stance signals that even the most well-intentioned open release carries unacceptable risks.

Anthropic draws a clear line: releasing model weights without restrictions (the 'open-weight' model) is too dangerous. Unlike closed APIs, weights can be fine-tuned, stripped of safeguards, and deployed at scale by malicious actors. The company points to examples where open models have already been used for harmful purposes, and argues that current safety evaluations are insufficient to guarantee responsible use.

But here's where it gets interesting: Anthropic doesn't call for a total ban. Instead, it endorses a middle path—what some call 'open-but-gated' or 'responsible release.' That means providing access to researchers and auditors under terms of use, while keeping weights out of the hands of bad actors. It's a pragmatic compromise that balances innovation with safety.

Critics will cry gatekeeping. But Anthropic's argument is compelling: the stakes are simply too high to leave weight distribution to the whims of open-source ideology. As the AI race accelerates, this position could become the new industry standard—if others have the courage to follow.

Bottom line: Anthropic's position isn't just a policy statement; it's a challenge to the entire AI community. Either we move toward controlled release and robust safety testing, or we accept that open weights will be exploited. There's no middle ground on catastrophic risk.