📰 Key Summary

Anthropic is taking a staged, controlled rollout approach with its most powerful models, arguing that once a new model goes public, attackers could benefit from it faster than defenders can. Delphi Digital researcher Su says that if a new model lets attackers scale up capabilities much faster than defenders, it ends up hurting the whole ecosystem, so a limited testing period helps shrink the potential “blast radius.” That said, he also notes that as other, more powerful models keep rolling out and spreading widely, Anthropic will face pressure to make its own models more accessible — and at that point, the key question becomes whether defenders can deploy these frontier models as effectively as attackers do. Michael Coates, Solana Foundation’s CISO (who joined in July), also backs the safety guardrails but argues the process should move faster for legitimate defenders to get access, simplifying verification and approval. Sean Cheetham of blockchain VC Blockchain Capital thinks the restrictions should ease up over the long run, since there are far more legitimate security researchers than small groups capable of pulling off sophisticated attacks — letting the good guys scale up their defenses would actually work in everyone’s favor. The article notes that Binance, the world’s largest crypto exchange with $137.8 billion in assets, currently doesn’t have access to frontier AI models. Crypto custodian Fireblocks said back in April it was applying for access to Anthropic’s Mythos program, at the time only using the public version of the model for penetration testing. Uniswap founder Hayden Adams criticized Fable 5 in June for being overly restrictive with cybersecurity-related prompts. The Ethereum Foundation said in July it’s running “coordinated AI agents” to hunt for vulnerabilities in its systems, though it didn’t disclose which model it’s using. Fintech company FIS (which partners with Circle on USDC payments) joined Anthropic’s gated Project Glasswing last month and got access.

💬 JudyAI Lab Take

In the original piece, Anthropic’s staged, controlled rollout for its most powerful models comes down to one reason: once a new model goes public, attackers might benefit from it faster than defenders can.

This gets at a core dilemma in AI safety design — if capability diffusion speeds up unevenly, with defenders falling behind attackers, the whole ecosystem pays the price. Delphi Digital’s Su points out that a limited testing window helps shrink the “blast radius,” but that’s a band-aid, not a cure — as stronger models keep shipping, the pressure to open up is only going to build. Solana’s CISO Coates wants faster access approvals for legitimate defenders, while Blockchain Capital’s Cheetham argues access should open up long-term, since there are way more good-faith security researchers out there than sophisticated attack groups. Binance still doesn’t have frontier model access, while FIS has already joined Anthropic’s Project Glasswing — which shows that access itself is becoming both a competitive edge and a source of security gaps.

For anyone building with AI, the takeaway here is that how model access gets allocated is already a security strategy you need to plan for proactively, not something to sort out after the fact.

📅 Original Article Info

🔗 Further Reading