Crypto Firms Seek Frontier AI Access as Only a Few Get In

Crypto security teams are facing a new imbalance: while AI model developers are restricting their most capable cyber-related systems, only a small number of crypto firms appear to have gained early access to those “frontier” tools.

Coinbase has said it secured access to Anthropic’s restricted Mythos model, and Zcash co-founder Zooko Wilcox has described how Anthropic used Mythos to audit the Zcash protocol at the request of Shielded Labs. Meanwhile, Binance’s chief security officer Jimmy Su told Cointelegraph that the exchange has been trying to make progress but has not obtained the most advanced frontier model.

Key takeaways

• Only select crypto companies have reportedly received early access to restricted frontier AI models used for cybersecurity work.
• Executives argue that gating advanced models may be necessary at first, but maintaining restrictions could become harder to justify as capabilities converge with public releases.
• Uneven access may widen the security gap between defenders and attackers, particularly as AI-assisted exploit workflows reportedly speed up.
• Some crypto-adjacent organizations, such as those embedded in critical infrastructure or security tooling, have also joined gated programs.

Why restricted “frontier” models are hard to distribute

The core issue is not whether AI can help security—many teams already use mainstream models for testing and review—but whether defenders get access to the most cyber-capable systems under developer guardrails.

Anthropic has stated that Mythos 5 shares the same underlying model as its publicly available Fable 5, but operates without safeguards that limit sensitive cybersecurity use. OpenAI is described as running a similar tiered approach, with a “Trusted Access for Cyber” pathway for verified defenders and a more permissive cyber-oriented version reserved for a smaller group conducting authorized penetration testing.

Crypto security executives interviewed by Cointelegraph suggested this kind of restricted rollout is likely warranted initially. However, they also highlighted a growing tension: once publicly available models begin to approach the same practical capabilities, continuous gating may become harder to defend—especially if attackers can leverage comparable tools from elsewhere.

Crypto executives push for faster verification pathways

Jimmy Su said Anthropic’s controlled release can be responsible because attackers may benefit from newly released capabilities sooner than defenders. In his framing, limiting early access can reduce the “blast radius,” at least during an initial testing period.

Solana Foundation’s chief information security officer Michael Coates supported guardrails but argued that “legitimate defenders” need a faster route to the models. He said the process should streamline verification and acceptance programs so security teams can use the best available systems to match the pace of exploitation.

Blockchain Capital’s Sean Cheetham expressed a similar long-term view. While restrictions can help avoid immediate misuse, broader availability could ultimately benefit defense because the population of legitimate security researchers is typically far larger than the small groups able to run highly sophisticated attacks. That scale dynamic—more defenders than adversaries—may flip the risk calculus over time.

Who has access—and who appears to be waiting

Despite being the world’s largest exchange by daily trading volume, Binance has not reportedly secured access to Mythos, according to Su. The exchange’s scale underscores the potential operational impact: Binance holds substantial assets on its platform, and a lack of frontier defensive tooling could leave major ecosystems to rely on less capable alternatives.

Beyond exchanges, other organizations have taken different approaches. Cointelegraph previously reported that Fireblocks, which provides custody and security services at large scale, sought access to Mythos but at the time relied on Anthropic’s publicly available model for pentesting. Cointelegraph also cited that Uniswap founder Hayden Adams criticized the safeguards on Fable 5 around cybersecurity prompts earlier this year.

The Ethereum Foundation has said it has been using “coordinated AI agents” to identify bugs across its systems, without disclosing which models were used. Cointelegraph reached out to the Ethereum Foundation, Fireblocks, and Uniswap to confirm whether they had received access to frontier restricted models since then.

Some crypto-adjacent companies, however, have moved further into gated programs. FIS—an infrastructure provider that partnered with Circle for USDC payments functionality last year—reportedly joined Anthropic’s Project Glasswing last month. Project Glasswing is described as Anthropic’s gated program for vetted cyber defenders and organizations responsible for critical software infrastructure to access restricted Mythos models.

HackerOne, which supports bug bounty and security testing for major organizations including crypto exchanges, also said it joined Project Glasswing, though its testing is confined to its own infrastructure rather than customer programs. Separately, Cointelegraph reached out to OpenAI and Anthropic to ask how many crypto firms had received access to restricted models.

AI-assisted attacks are reported to be accelerating

The access gap matters because defenders are not operating in a static threat environment. Cointelegraph reported that Boltz, a Bitcoin swap service, chose to halt its non-custodial bridge after observing a steady rise in AI-assisted hacking attempts over the prior few months. Boltz said attackers are now iterating faster than a team its size can find and patch, pointing to a practical pressure on incident response and code review cycles.

Other security events also suggest attackers are applying automation to find real weaknesses. Cointelegraph reported that Coinkite disclosed a vulnerability affecting some Coldcard devices, where a flaw in wallet seed generation produced less randomness than expected. Coinkite speculated that the attacker could have used AI to review earlier firmware versions to locate and exploit the weakness—even though the company had used what it described as one of the best available AI models to review its code in the weeks before.

Taken together, these reports support a broader concern: even strong internal testing using public AI tools may not be enough if adversaries deploy higher-end capabilities and iterate faster than teams can remediate.

What to watch next for crypto security

The immediate question is whether restricted model access will widen beyond early adopters and whether developers can design guardrails that protect the ecosystem without bottlenecking legitimate defenders. As AI capabilities diffuse—through both public releases and competing models—crypto teams will likely watch not only for new access announcements, but also for changes in exploitation tempo and patch turnaround times across major platforms.