Key Highlights

  • Anthropic released Claude Fable 5, its most capable public AI model to date, alongside a restricted variant called Claude Mythos 5 available only to vetted cybersecurity researchers.

  • DeFi protocols have lost more than $840 million to hacks in the first five months of 2026, with April alone accounting for over $600 million.

  • Experts warn that AI's core contribution to attackers is not the invention of new hacking techniques but machine-speed execution of reconnaissance work, scanning codebases.

Anthropic launched Claude Fable 5 this week, describing it as its most powerful public model yet and releasing it in two forms. The standard version is available broadly, with safety filters designed to block its most dangerous uses. A second, less-restricted variant called Claude Mythos 5 is distributed only to vetted users in cybersecurity and critical infrastructure, where its ability to identify and chain zero-day vulnerabilities into working attacks can be used defensively. Anthropic said its safety filters reroute high-risk requests to a weaker model in fewer than 5% of sessions, and that over 1,000 hours of external bug-bounty testing found no universal bypass, though the company acknowledged that well-funded adversaries will keep trying.

The release lands at a difficult moment for decentralized finance. Data tracking protocol losses shows DeFi platforms have shed more than $840 million to hacks so far this year, with April alone marking the worst single month on record at over $600 million. The two largest incidents illustrate the kind of threat AI is best positioned to amplify. A North Korea-linked group spent six months executing a social-engineering campaign to gain administrator access before draining approximately $285 million from Drift Protocol. A separate single-verifier flaw allowed roughly $292 million to be siphoned from Kelp DAO. Neither attack required a novel smart-contract exploit. A third major incident hit the same week Fable 5 launched, when Humanity Protocol lost over $30 million after a hacker obtained three of six private keys stored on a single employee's laptop.

Security professionals say this pattern reveals where AI actually changes the threat landscape. Charles Guillemet, chief technology officer at hardware-wallet maker Ledger, noted that AI guardrails raise friction but are not a reliable control against a determined adversary. The more significant shift is speed: a reasoning model can read public repositories, compare software versions, map every configuration and enumerate every misconfiguration at machine speed, compressing the scouting phase of an attack from weeks to hours. The underlying techniques remain the same ones that have driven DeFi losses all year: social engineering, bad signing flows, exposed keys and human error. AI did not create those vulnerabilities. It made finding them dramatically faster.

The same capabilities work in defenders' favor as well. DeFi protocol Pendle said it has used Anthropic's models since an earlier version to map its codebase, stress-test contracts and catch bugs before deployment. The developers noted that smart contracts are actually among the easier targets to audit manually, since they are short and have a limited number of entry points. The more dangerous attack surface is the human and operational layer surrounding a protocol: the developers, key holders, signers and administrators whose mistakes create the openings that bug bounty researchers and, increasingly, AI-assisted attackers have consistently found and exploited in 2026.