Cryptocurrency companies are still seeking cutting-edge AI access, only a few have been approved

Cryptographic security leaders face a growing imbalance: Attackers benefit from cutting-edge AI assistance capabilities, while many large crypto companies still lack direct access to the strongest "cutting-edge" network models for security testing and code hardening

In June this year, Coinbase said it had obtained access to Anthropic's restricted model Mythos. Zcash founder Zooko Wilcox said Anthropic used Mythos to assist in auditing the Zcash protocol at the request of Shielded Labs. However, other large companies, including Binance, have publicly failed to obtain similar permissions-highlighting a new "security gap" in the industry.

Core Points

Large encryption companies have uneven access to restricted cutting-edge network models such as Anthropic's Mythos, resulting in some companies having fewer defense tools than others. Executives believe initial limitations on advanced models are necessary because attackers may master new capabilities faster than defenders. As public models close the capability gap, industry pressure may push defenders to gain broader access to restricted tools. Recent incidents involving AI-assisted attacks and wallet/bridge security show why accelerating defense iterations has become critical.

Why "cutting-edge" network models are unevenly obtained

Anthropic's Mythos is positioned as a restricted version of a model that is also related to a more public version-its developers say the underlying foundation is similar, but Mythos removes some safeguards that restrict sensitive network security efforts. OpenAI also describes a similar approach that provides different levels of capabilities based on user type and intended use, including providing "trusted access to network security" for verified defenders.

Jimmy Su, chief security officer of Binance, told Cointelegraph that the gap remains significant. According to Su, Binance has been working hard to acquire such tools, including dialogue with other crypto exchanges and investors, but has not yet acquired state-of-the-art models like Mythos.

This uneven promotion is important because the cyber threat landscape changes faster than traditional security review cycles. In ecosystems where vulnerabilities can be exploited quickly (often assisted by automation or semi-automation), fewer defense options mean vulnerabilities are discovered and repaired more slowly.

Executives support initial restrictions-but question long-term rationality

Cryptographic security executives interviewed by Cointelegraph said there is likely to be a reasonable need to initially restrict access to cutting-edge network models. The reason is simple: If the new model enhances attackers faster than defenders, it could increase the ecosystem's exposure to risk before the security community catches up.

Su believes that controlling the scope of release can narrow the "explosion radius", especially in the early stages. However, he also pointed out that this rationale will change as competitive models become stronger and more widely available. In this case, pressure turns to model developers to expand access-the key question is whether defenders can effectively use these tools at the same speed as attackers.

Michael Coates, chief information security officer of the Solana Foundation, echoed this contradiction. Coates supports security safeguards but says verification and acceptance paths can slow down legitimate defensive use. In his view, defenders need a more streamlined process to give teams that can evaluate code and identify problems access to advanced models before the code is leveraged.

Sean Cheetham of Blockchain Capital also favors an eventual opening. He believes that while malicious actors are highly skilled, the broader group of security researchers is usually larger. If "good people" can expand defenses, broader access may ultimately strengthen the ecosystem more than helping attackers.

Large exchanges and builders are still waiting, while some crypto-related companies have been allowed to enter

Binance's access is remarkable because of its size: DefiLlama data cited by Cointelegraph shows Binance's total assets reach US$137.8 billion. Despite its size, Su said the exchange has not yet reached the highest level of access to cutting-edge network models.

Cointelegraph also reported periodic signals of gaining authority within the industry. Fireblocks, a large crypto custodian, said in April that it had sought access to Mythos. At the time, it relied on Anthropic's public model for penetration testing rather than restricted access. Uniswap founder Hayden Adams also criticized the safeguards attached to Fable 5-related cybersecurity tips.

In addition, the Ethereum Foundation said in July that it had been running "coordinated AI agents" to find system vulnerabilities, but did not specify which model it used. Cointelegraph contacted the Ethereum Foundation, Fireblocks and Uniswap to confirm whether they had gained access to restricted cutting-edge models since earlier announcements.

Although many crypto companies still seem to be waiting, some crypto-related organizations have already taken the lead. FIS, which provides technology to banks and partnered with Circle to launch USDC payments last year, said it joined Anthropic's Project Glasswing program last month. The program aims to provide proven cyber defenders and critical software infrastructure organizations with early access to the restricted Mythos model. HackerOne, known for its vulnerability rewards and security testing, also said it has joined Project Glasswing. Cointelegraph pointed out that in this case, testing was limited to HackerOne's own infrastructure and not extended to customer projects.

Cointelegraph contacted OpenAI and Anthropic to ask how many crypto companies had been allowed access to restricted models, but these responses were not included in the text of the article provided.

Why risk escalates: AI-assisted attacks and faster attacker iteration

The background of the access battle is frequent security incidents, which organizations attribute to faster attack utilization cycles. On Monday, Bitcoin exchange service Boltz said it had suspended its unmanaged bridging after observing a steady increase in AI-assisted hacking attempts over the past few months. Boltz's statement argued that attackers can iterate faster than small security teams and discover vulnerabilities to patch them, Cointelegraph reported.

Hardware wallet maker Coinkite reported last week that some of its Coldcard devices were being exploited due to wallet seed generation flaws. Coinkite said seed generation was less random than expected and speculated that attackers may have used AI to check previous firmware versions to identify and exploit the vulnerability-although the company said it only used "one of the best AI models available at the time" to review its code a few weeks ago.

These examples highlight a key practical issue: Even if a defender uses advanced tools, a rhythm mismatch (the speed at which the attacker adapts and the speed at which the defender verifies and fixes) can still determine the outcome. Obtaining limited models may be only part of the answer; process, test rigor, and deployment speed remain critical to reducing real-world risks.

Next, readers should focus on whether cutting-edge model providers are expanding defender access beyond current limited channels, and whether security teams can demonstrate that broader availability improves results rather than accelerates attacks. Who can test with the most powerful tools and how quickly they can patch vulnerabilities-a gap that could become one of the most defining operational gaps in the realm of cryptographic security.