Anthropic launched Claude Fable 5 on Tuesday, the first public model in its Mythos class, alongside a restricted version called Claude Mythos 5 limited to vetted users in cybersecurity and critical infrastructure. The company claims the model can find and chain zero-day vulnerabilities—previously unknown software flaws—and help turn them into working exploits. Despite safety guardrails that route high-risk requests to a weaker fallback model (triggered in fewer than 5% of sessions), and over 1,000 hours of external bug-bounty testing, Anthropic openly admits the system is unlikely to be foolproof against determined, well-funded attackers.
Speed, Not Novelty, Is the Game-Changer
Charles Guillemet, CTO of hardware-wallet maker Ledger, said the biggest AI-driven shift is not inventing new attack types but compressing the time to create them. A reasoning model can “diff every commit, grep every config, and enumerate every misconfiguration at machine speed,” referencing software development steps. Crypto is especially exposed because software failure can become a financial loss almost immediately. Anthropic itself noted in a blog post: “The uplift from Mythos-level capabilities is valuable to many adversaries ... we therefore expect them to be motivated to try to circumvent our safety measures.”
DeFi Losses Top $840M in 2025, but Social Engineering Still Dominates
DeFi protocols lost more than $840 million to hacks in the first five months of 2026, per DefiLlama, with April alone accounting for over $600 million—the worst month on record. Yet the two biggest incidents did not involve smart-contract exploits AI could engineer: a North Korea-linked group drained ~$285 million from Drift Protocol after a six-month social-engineering campaign, and an attacker exploited a single-verifier flaw to siphon ~$292 million from Kelp DAO. On Tuesday, Humanity Protocol lost over $30 million to a private-key compromise—a hacker accessed three of six keys stored on a single employee's laptop.
Old Weaknesses, Accelerated
Guillemet noted that these exploits remain rooted in social engineering, bad signing flows, exposed keys and human error. A model like Fable 5 doesn't need to output a finished exploit: it can read public repos, compare old software versions, summarize audit reports, and draft convincing messages targeting small operational mistakes humans miss. Defenders must secure every key path, dependency, signing flow and privileged account. With AI accelerating the scouting phase, the final signing step becomes critical—private keys should live where a compromised laptop can't reach, and users need a trusted screen showing exactly what they approve.
Crypto's next billion-dollar hack may not come from a brilliant contract bug. It will be a race between superhuman speed and hardened operational security.