Anthropic released Claude Fable 5, its most powerful AI model, with a restricted version (Mythos 5) for vetted cybersecurity users. The model offers superhuman speed in finding misconfigurations, social engineering exploits, and flawed signing flows, but safety filters may not stop determined adversaries. DeFi has lost over $840 million to hacks in 2025, mainly from human error and operational failures, not smart-contract bugs. Experts warn AI accelerates scouting but doesn't create new hack types. Hardware wallets with trusted displays and Clear Signing are recommended as a defense. Pendle uses the same AI defensively to audit contracts.

Key facts

• Claude Fable 5 offers stronger reasoning and coding, with Mythos 5 restricted to vetted cybersecurity users.
• DeFi hacks in 2025 total over $840 million, mostly from social engineering and human error.
• AI speeds up finding misconfigurations and exploiting weak points but doesn't invent new hack types.
• Anthropic's safety filters may block <5% of dangerous requests; determined attackers may bypass them.
• Pendle uses Anthropic's models defensively to audit smart contracts and catch bugs early.