Major crypto companies are trying to obtain the most capable AI models for vulnerability research, but access remains uneven. Coinbase has described work with Anthropic's restricted Mythos model, and the technology was used in a Zcash protocol audit. Binance and other firms are still seeking comparable capabilities, according to security executives. Model developers restrict systems with strong cyber abilities because they could also accelerate exploitation.

The safety logic is understandable, yet crypto faces an uncomfortable asymmetry. An attacker does not need formal access to one restricted model. They can combine open models, stolen credentials, multiple agents and automated testing across thousands of targets. A regulated exchange, by contrast, must document its use, satisfy provider requirements and wait for approval.

That creates a new form of infrastructure inequality. Organizations considered mature enough by AI laboratories can analyze code and simulate attacks faster. Small protocols remain on weaker tools even though they have fewer engineers and less round-the-clock monitoring. Security could split between privileged companies with frontier AI and a long tail that learns about a new technique only after an exploit.

The answer is not simply releasing every model. The industry needs controlled sandboxes, logged queries, independent access programs and shared testing environments that do not require distributing model weights. Watch the criteria used by Anthropic and OpenAI, new security consortiums and whether audit firms can offer powerful models as a monitored service. The central question is whether frontier AI becomes common defensive infrastructure or another advantage reserved for the largest platforms. In a market of irreversible transactions, delayed defensive access has a measurable financial cost.