Crypto Security Gap Widens as Elite AI Access Remains Restricted

Woofun AI reports that a stark security divide has emerged within the cryptocurrency industry, driven by the unequal distribution of access to high-capability artificial intelligence models. While entities such as Coinbase and Zcash have successfully secured entry to restricted frontier tools, including Anthropic’s Mythos model, other major players like Binance remain excluded, creating a significant vulnerability gap. This disparity was highlighted by Jimmy Su, Binance’s chief security officer, who confirmed that despite extensive efforts, his organization has not yet obtained access to the most advanced AI models available for cybersecurity defense.

The specific mechanics of this exclusion became apparent in June, when Coinbase announced it had secured access to Anthropic’s restricted Mythos model. Simultaneously, Zooko Wilcox of Zcash revealed that Anthropic had utilized the same model to audit the Zcash protocol at the request of Shielded Labs. In contrast, Jimmy Su told Cointelegraph that Mythos remains one of the few advanced frontier models not yet made available to the broader crypto sector. Su noted that Binance has attempted to make inroads by engaging with other exchanges and investors, but these efforts have not resulted in access to the most frontier AI capabilities.

This uneven access creates a profound security divide in an industry where exploits can put billions of dollars at risk. Model developers like Anthropic and OpenAI have chosen to restrict their most cyber-capable models from the general public, citing safety concerns.

However, this restriction raises fears that increasingly powerful open-source alternatives will allow attackers to outpace defenders. Crypto firms are thus left to contend with sophisticated AI-assisted attacks without possessing the most capable tools necessary to defend against them effectively.

Anthropic’s approach to model architecture illustrates this tension between capability and control. The company states that Mythos 5 uses the same underlying model as its publicly available Fable 5, but it operates without the safeguards that restrict sensitive cybersecurity work. This distinction allows Mythos to perform tasks that Fable 5 is programmed to avoid, making it a critical tool for advanced security analysis. The removal of these safeguards is intended to enable deeper inspection of code and infrastructure, but it also necessitates strict access controls to prevent misuse.

Woofun AI data shows that OpenAI employs a similar tiered access system to manage the risks associated with its most powerful models. Verified defenders can utilize GPT-5.5 with "Trusted Access for Cyber", which provides enhanced capabilities for security professionals.

However, the more permissive GPT-5.5-Cyber model is reserved for a smaller, highly vetted group conducting authorized penetration testing. This structure aims to balance the need for robust defensive tools with the imperative to prevent these tools from falling into malicious hands.

The industry debate over gatekeeping versus availability centers on the potential "blast radius" of unrestricted access. Jimmy Su argued that Anthropic’s controlled rollout is a responsible approach, as newly released models may benefit attackers faster than defenders. He stated, "If it enhances the attacker much faster than the defender, then it actually is harming the ecosystem." Su suggested that a limited testing period could reduce the potential damage, but noted that this calculation changes as competing models become more powerful and widely available.

Calls for streamlined defender access have grown louder among security executives. Michael Coates, Solana Foundation’s chief information security officer, who joined the foundation in July, supported safeguards but argued that legitimate defenders need a faster route to them. He emphasized the need to streamline verification programs to ensure the best models are in the hands of defenders. Sean Cheetham of Blockchain Capital also advocated for eventually opening up restrictions, arguing that broader availability could favor defenders since legitimate security researchers greatly outnumber sophisticated attack groups.

Major players denied access or criticizing restrictions highlight the scale of the issue. Binance, despite being the biggest crypto exchange by daily trading volume and holding $137.8 billion in assets according to DefiLlama, remains without access. Fireblocks, a crypto custodian securing trillions in assets, said in April it had sought access to Mythos but was only using Anthropic’s publicly available model for pentesting, according to The Information. Uniswap founder Hayden Adams slammed Fable 5’s safeguards in June, criticizing the restrictions on prompts relating to cybersecurity.

Other entities have found varying degrees of success in gaining access. The Ethereum Foundation stated in July that it has been running "coordinated AI agents" to find bugs across its systems, though it did not disclose which models were used.

Meanwhile, FIS, which partnered with Circle in July last year to enable USDC payments, joined Project Glasswing last month. This gated program by Anthropic provides vetted cyber defenders early access to restricted Mythos models. HackerOne also joined Project Glasswing, though its testing is confined to its own infrastructure, not its customers’ programs.

Recent exploits underscore the urgency of this access gap. On Monday, Bitcoin swap service Boltz halted its non-custodial bridge after observing a steady rise in AI-assisted exploits, noting that attackers now iterate faster than their team can patch. Last week, Bitcoin hardware wallet company Coinkite reported that its Coldcard devices were exploited due to a flaw in wallet seed generation. Coinkite speculated that the attacker used AI to review previous firmware versions, despite the company having used "one of the best available AI models" to review its code just weeks prior.