Crypto’s next billion-dollar hacker may move at superhuman speed
Anthropic’s new Claude Fable 5 could speed up crypto attacks even with safety filters. Experts say the bigger danger is not new hacks, but faster exploitation of human mistakes.
Intelligence analysis by GPT-5.4 Mini
Anthropic released Claude Fable 5 with stronger reasoning and coding, while limiting its most capable variant to vetted security users. In crypto, where most big losses come from operational failures and social engineering, that speed-up could make attackers far more effective.
A super-smart computer helper can now search for weak spots very fast, like a burglar who can check every window in seconds. That does not create new locks to break, but it can help thieves find open doors much faster.
Analysis
What Anthropic shipped
Anthropic introduced Claude Fable 5 as its newest public model, with stronger reasoning and coding abilities but filters meant to block dangerous uses. A more powerful version, Claude Mythos 5, is restricted to vetted cybersecurity and critical infrastructure users. The company says suspicious requests can be routed to a weaker fallback model, Claude Opus 4.8, and says outside testing did not find a universal bypass.
Why crypto is exposed
The article argues that AI is less likely to invent entirely new crypto exploits than to make existing ones much faster to find and execute. Charles Guillemet of Ledger said current guardrails only add friction and are not a reliable defense against a determined attacker. A model like this can inspect public repositories, compare software changes, read audit reports, and draft convincing messages that help attackers find misconfigurations, exposed keys, and weak signing flows.
The losses behind the warning
DeFi has already lost more than $840 million to hacks in the first five months of the year, according to DefiLlama data. April alone saw more than $600 million in losses. The biggest incidents were not simple smart-contract bugs: one attack on Drift Protocol followed a long social-engineering campaign that won admin access, another exploited a single-verifier flaw at Kelp DAO, and Humanity Protocol lost more than $30 million after a private-key compromise.
The practical takeaway
The piece’s core point is that AI can accelerate the scouting phase of an attack. That makes the final approval step even more important: keys need to be isolated, signing flows need to be clear, and users need trusted hardware and screens that show exactly what they are approving.
Key points
- Anthropic released Claude Fable 5 with stronger reasoning and coding, plus safety filters for dangerous requests.
- A more powerful Claude Mythos 5 is limited to vetted cybersecurity and critical infrastructure users.
- Experts say AI may not create brand-new crypto hacks, but it can speed up finding misconfigurations and weak points.
- DeFi has lost more than $840 million to hacks in the first five months of the year, according to DefiLlama data.
- The biggest losses cited in the article came from social engineering, key compromise, and signing or access failures.
If the stronger model and its safety filters hold up, it could also help defenders test systems faster and find weaknesses before attackers do. The article suggests security teams and bug bounty testing are already being used to probe the safeguards.
If determined attackers find a way around the filters, they could use the model to speed up phishing, key theft, and other human-centered attacks. Because many of the biggest crypto losses came from social engineering and operational mistakes, faster attack workflows could translate into larger losses.
