Crypto Firms Still Seek Frontier AI Access, With Only Few Approved

Crypto security leaders are grappling with a growing mismatch: attackers increasingly benefit from cutting-edge, AI-assisted capabilities, while many major crypto firms still lack direct access to the most powerful “frontier” cyber models used for security testing and code hardening.

In June, Coinbase said it had secured access to Anthropic’s restricted Mythos model, and Zcash founder Zooko Wilcox said Anthropic used Mythos to help audit the Zcash protocol at the request of Shielded Labs. Yet other large players, including Binance, have publicly indicated they have not been able to obtain similar access—highlighting an emerging “security divide” across the industry.

Key takeaways

• Major crypto firms have uneven access to restricted frontier cyber models like Anthropic’s Mythos, leaving some with fewer defensive tools than others.
• Executives argue gating advanced models is initially necessary because attackers may adopt new capabilities faster than defenders.
• As publicly available models close the capability gap, industry pressure is likely to increase for wider defender access to restricted tools.
• Recent incidents involving AI-assisted exploitation and wallet/bridge security show why faster defensive iteration is becoming critical.

Why access to “frontier” cyber models is uneven

Anthropic’s Mythos is positioned as a restricted version of a model also related to a more public offering—its developers say the underlying base is similar, but that Mythos removes certain safeguards that limit sensitive cybersecurity work. OpenAI has described a comparable approach, offering different tiers of capability depending on the type of user and intended use, including “Trusted Access for Cyber” for verified defenders.

Binance chief security officer Jimmy Su told Cointelegraph that the gap remains meaningful. According to Su, Binance has been trying to make progress on obtaining such tools, including conversations with other crypto exchanges and investors, but it has not obtained the most advanced model like Mythos.

This uneven rollout matters because the cyber threat landscape is moving faster than traditional security review cycles. In an ecosystem where vulnerabilities can be exploited rapidly—often with automated or semi-automated assistance—having fewer defensive options can translate into slower discovery and patching.

Executives back initial gating—then question the long-term rationale

Crypto security executives interviewed by Cointelegraph said there is likely a legitimate need to restrict initial access to frontier cyber models. The reasoning is straightforward: if a new model enhances attackers more quickly than defenders, it can increase the ecosystem’s exposure before the security community catches up.

Su suggested that controlled rollout can reduce the “blast radius,” especially early on. However, he also argued that the justification changes as competing models become more powerful and more widely available. In that scenario, the pressure shifts to model developers to broaden access—and the key question becomes whether defenders can use the tools effectively at the same pace attackers do.

Michael Coates, chief information security officer at the Solana Foundation, echoed the tension. Coates supported safeguards but said verification and acceptance pathways can slow down legitimate defensive usage. In his view, defenders need a more streamlined process to get advanced models into the hands of teams that can evaluate code and identify issues before they are exploited.

Blockchain Capital’s Sean Cheetham also leaned toward eventual opening. He argued that while malicious actors are skilled, the broader pool of security researchers tends to be larger. If “good people” can scale their defensive work, broader access could ultimately strengthen the ecosystem more than it helps attackers.

Large exchanges and builders still waiting, while some crypto-adjacent firms got in

Binance’s access situation is notable given its scale: DefiLlama data cited by Cointelegraph places Binance’s total assets at $137.8 billion. Despite that footprint, Su indicated the exchange had not reached the highest tier of frontier cyber-model access.

Cointelegraph also reported additional signals of staggered access within the sector. Fireblocks, a major crypto custodian, said in April it had sought access to Mythos. At that time, it relied on Anthropic’s publicly available model for penetration testing rather than restricted access. Uniswap founder Hayden Adams similarly criticized safeguards tied to cybersecurity prompts in relation to Fable 5.

Separately, the Ethereum Foundation said in July that it has been running “coordinated AI agents” to find bugs across its systems but did not specify which models were being used. Cointelegraph reached out to the Ethereum Foundation, Fireblocks, and Uniswap to confirm whether they had received access to restricted frontier models since those earlier statements.

While many crypto players appear to be waiting, some crypto-adjacent organizations have moved ahead. FIS, which provides technology to banks and partnered with Circle last year for USDC payments, said it joined Anthropic’s Project Glasswing program last month. Project Glasswing is designed as a gated channel for vetted cyber defenders and critical software infrastructure organizations to obtain early access to restricted Mythos models.

HackerOne—known for bug bounty and security testing—also said it joined Project Glasswing. Cointelegraph notes that, in this case, testing is limited to HackerOne’s own infrastructure rather than being extended to customer programs.

Cointelegraph reached out to OpenAI and Anthropic for details on how many crypto companies have been granted access to restricted models, but those responses are not included in the article text provided.

Why the stakes are rising: AI-assisted exploitation and faster attacker iteration

The access debate is playing out against a backdrop of security incidents that organizations link to faster exploitation cycles. On Monday, Bitcoin swap service Boltz said it paused its non-custodial bridge after it observed a steady rise in AI-assisted hacking attempts over the past few months. Boltz’s statement, as reported by Cointelegraph, argued that the pattern is that attackers can iterate faster than a smaller security team can find and patch issues.

Hardware wallet maker Coinkite reported last week that some of its Coldcard devices were exploited due to a flaw in wallet seed generation. Coinkite indicated the randomness of the seed generation was less than expected and speculated that the attacker may have used AI to examine previous firmware versions to identify and exploit the weakness—even though the company said it had used “one of the best available AI models” to review its code only weeks earlier.

These examples underscore a key practical problem: even when defenders use advanced tools, the cadence mismatch—how quickly attackers can adapt and how quickly defenders can verify fixes—can still drive outcomes. Access to restricted models may be only part of the answer; process, testing rigor, and deployment speed remain central to reducing real-world risk.

Next, readers should watch whether frontier-model providers expand defender access beyond the current limited pipelines and whether security teams can demonstrate that broader availability improves outcomes rather than accelerating exploitation. The gap between who can test with the most capable tools—and how quickly they can patch—may become one of the defining operational fault lines in crypto security.