Zooko Wilcox Su, a representative associated with the Solana Foundation, said that Anthropic's decision to gradually roll out its Claude Mythos model is a responsible strategy. He emphasized that newly launched models might give attackers a quick edge over defenders, potentially damaging the broader technology environment. However, Su pointed out that if stronger models become more accessible over time, Anthropic could face increasing demands to expand availability. He stressed that the balance shifts when competitors release models that are more powerful and widely adopted.
Michael Coates, who serves as the chief information security officer at the Solana Foundation and joined the group in July, supports protective measures for advanced AI systems. He expressed his full backing for these safeguards but argued that the process for vetting and granting access should be streamlined for genuine defenders. Coates highlighted the need to ensure that the most advanced AI models reach defenders before attackers can take advantage of their capabilities. He noted that attackers already have tools powerful enough to exploit security weaknesses.
Binance and Fireblocks left out of the loop
Despite being the top cryptocurrency exchange in terms of daily trading volume, Binance has not received access to Anthropic's restricted AI model. The company manages assets worth $137.8 billion, as reported by DefiLlama. According to The Information, Binance has been relying on Anthropic's publicly available model for penetration testing purposes. Crypto custodian Fireblocks, which is responsible for securing trillions in assets annually, also sought access to the Mythos model in April. At that time, Fireblocks used only the publicly accessible version for pentesting.
Hayden Adams, the founder of Uniswap, criticized Fable 5’s security measures in June, particularly those that limit prompts related to cybersecurity. Meanwhile, in July, the Ethereum Foundation announced it had begun using coordinated AI agents to identify bugs within its systems. However, the foundation did not reveal the specific models it was employing in this effort.
Crypto-related companies granted access to advanced models
FIS, a technology provider to banks, recently joined Project Glasswing after forming a partnership with Circle in July of the previous year to offer cross-border payments in USDC. Project Glasswing is a restricted initiative by Anthropic that provides vetted defenders and key infrastructure organizations with early access to its advanced AI models. HackerOne, a company offering bug-bounty and security testing services to major crypto exchanges, also became part of the program. However, HackerOne limited its testing to its own infrastructure and not that of its clients. Cointelegraph inquired with OpenAI and Anthropic about how many crypto-related companies had been granted access to restricted models but had not received responses at the time of publication.
AI-assisted attacks target crypto vulnerabilities
In a recent incident, a Bitcoin hardware wallet company named Coinkite reported that some of its Coldcard devices had been compromised due to a flaw in the wallet's seed generation. The issue was attributed to a lack of randomness in the process. The company speculated that an attacker had used AI tools to examine earlier versions of the firmware and identify exploitable weaknesses. Despite using one of the best available AI models to scan its code just weeks before the incident, the flaw still went undetected.
Sean Cheetham from Blockchain Capital suggested that giving defenders broader access to advanced AI models could allow them to scale their security efforts faster than attackers can launch new exploits. He argued that legitimate security researchers vastly outnumber the small groups of attackers who carry out sophisticated cyber operations. According to Cheetham, by enabling defenders to leverage AI tools on a larger scale, the overall security landscape could improve significantly.
The response status from these entities was unclear at the time of publication. Meanwhile, Bitcoin swap service Boltz announced in recent days that it has decided to halt its non-custodial bridge due to a steady rise in AI-assisted attacks over the past few months. The company said the trend was clear: attackers now evolve and exploit vulnerabilities faster than its team can identify and fix them.
Cointelegraph remains dedicated to delivering independent, transparent journalism, as outlined in its Editorial Policy. The article is intended to offer accurate and timely information to readers, who are encouraged to verify facts independently.