AI’s New Tool Could Speed Up Crypto Hacks, But Also Help Stop Them

Anthropic’s latest AI model, Claude Fable 5, is a powerful new tool that can help people write code faster and think more clearly. The company released two versions: a public one that blocks dangerous requests, and a restricted version called Claude Mythos 5 for experts in security. The restricted model can spot hidden software bugs that no one knows about yet, and it can chain them into real attacks. Anthropic says the system watches for risky requests and sends them to a weaker model, Claude Opus 4. 8, in less than five percent of cases. Security teams have tested it for over a thousand hours and found no easy way to break it, but the company admits determined attackers may still try. Even with guardrails, AI can find weaknesses very quickly. A CTO at a hardware wallet company said the controls add friction but do not stop a determined attacker. The real threat is how fast AI can scan code, find misconfigurations and draft convincing phishing messages.Crypto projects lose a lot of money to hacks each year, often through social engineering or human error rather than smart‑contract bugs. Recent attacks drained millions from DeFi protocols by tricking admins or exploiting simple signing mistakes. Because AI can read public code, compare versions and summarize audits, it can help attackers spot small operational mistakes that humans overlook. That means defenders must secure every key, signing flow and privileged account. The best defense is to keep private keys in a trusted hardware device that shows what the user is approving. On the defensive side, some projects already use AI to map their code and test contracts before they go live. One DeFi protocol said the tool finds bugs early and makes code cleaner. The next big hack may not be a new type of exploit, but the same old tricks—poisoned packages, fooled developers or bad signing flows—that crypto has faced for years. AI just makes finding and using those tricks faster.