Crypto firms seek access to restricted frontier AI models

Several major crypto firms are seeking access to restricted frontier AI models such as Anthropic’s Mythos. Only a small number of exchanges, custodians and vetted security teams currently have access.

Anthropic has given early access through Project Glasswing to organizations responsible for critical software infrastructure. Coinbase has confirmed access. Zooko Wilcox of Zcash said Mythos was used to audit the Zcash protocol at the request of Shielded Labs. Financial-services firm FIS and security company HackerOne joined the program; HackerOne’s testing has been limited to its own systems.

Binance’s chief security officer, Jimmy Su, has tried to gain entry but has not received the most advanced model. He warned that releasing new models too widely could help attackers faster than defenders: “If it enhances the attacker much faster than the defender, then it actually is harming the ecosystem.”

Michael Coates, chief information security officer at the Solana Foundation, urged faster verification for vetted defenders: “We need to make sure that the best models we can get are in the hands of defenders because attackers will have something capable enough.”

Sean Cheetham of Blockchain Capital noted that wider access could benefit defenders because legitimate researchers outnumber skilled attackers: “If good people can multiply their defense scale… you’re much better off just opening it up and allowing them to defend themselves.”

Several crypto services have reported AI-assisted attacks. Bitcoin swap service Boltz paused its non-custodial bridge after attackers moved faster than the team could patch. Hardware wallet maker Coinkite reported some Coldcard devices were compromised after an attacker exploited a flaw in seed generation and indicated AI may have been used to find regressions in prior firmware. Security teams have observed a rise in critical-severity vulnerabilities following previews of advanced models.

Anthropic says Mythos 5 shares its architecture with public Fable 5 but removes safeguards that limit cybersecurity tasks. Other developers use tiered access: verified defenders can use cyber-capable versions while more permissive variants remain tightly controlled.

Fireblocks has sought access to restricted models and continues to use public models for penetration testing. Uniswap founder Hayden Adams criticized safeguards in Fable 5 that block cybersecurity prompts. The Ethereum Foundation is running coordinated AI agents to search for bugs but has not disclosed which models it uses.

Model developers cite risk management for gating powerful models. Security leaders request faster verification and wider access for vetted defenders to test and secure software infrastructure.