Crypto firms chase restricted Anthropic Mythos AI access as attacks

Crypto firms are racing to secure Anthropic’s restricted Mythos AI model after reports that AI‑assisted attacks on Bitcoin bridges and hardware wallets have accelerated, prompting calls for faster defender access [1].

Limited access to frontier AI models

Anthropic’s “Project Glasswing” grants vetted security teams entry to Mythos 5, a version of the model that removes the safety filters present in the public Fable 5 and is capable of advanced cybersecurity tasks [1]. Coinbase has confirmed it holds access, while Zcash co‑founder Zooko Wilcox disclosed that Mythos was used to audit the Zcash protocol at Shielded Labs’ request [1]. Binance’s chief security officer Jimmy Su says the exchange has not yet obtained the most advanced model, warning that broader release could empower attackers faster than defenders [1].

Growing threat from AI‑enhanced attacks

Recent incidents underscore the urgency. Boltz, a Bitcoin swap service, paused its non‑custodial bridge after attackers leveraged AI to outpace its patching speed [1]. Likewise, hardware‑wallet maker Coinkite reported that a flaw in seed generation was likely discovered using AI to comb through prior firmware revisions [1]. Security teams have observed a surge in critical‑severity vulnerabilities coinciding with previews of these advanced models [1].

Calls for faster defender verification

Solana Foundation CISO Michael Coates and Blockchain Capital’s Sean Cheetham argue that the verification process for “verified defenders” should be accelerated, noting that the pool of legitimate security researchers far exceeds the small groups capable of running sophisticated attacks [2]. HackerOne has joined Project Glasswing but limits its testing to its own systems, while Fireblocks continues to rely on public AI models for penetration testing [2]. The tension remains: developers cite risk management for gating powerful models, yet defenders fear the gap will widen as public releases approach the capabilities of restricted versions [2].

What to watch

• Access announcements – Monitor statements from Anthropic and OpenAI on expanding “Trusted Access for Cyber” programs to additional crypto firms.
• AI‑assisted exploit frequency – Track reported incidents of AI‑driven hacks on crypto bridges and hardware wallets for changes in tempo.
• Model rollout timelines – Watch for any scheduled updates to Mythos 5 or similar frontier models that could shift the defender‑attacker balance.

The scramble for restricted AI tools highlights a pivotal inflection point: if defenders cannot match the speed of AI‑enhanced attackers, the security of crypto infrastructure may hinge on how quickly model developers broaden vetted access.