What happened
Anthropic launched Claude Fable 5 on Friday, a new frontier model that the company says includes powerful cyber-offense capabilities placed behind a stack of safety filters, CoinDesk reported Friday. The framing matters. Anthropic isn't pretending the model can't do harm. It's betting the filters will catch the prompts that try.
The pitch from Anthropic is that gated capability beats forced ignorance. Earlier-generation models played dumb when asked about exploit primitives. Fable 5 reportedly knows the answer and refuses to give it. Refusal is a brittle defense, and that's the part security teams are reading twice.
The CoinDesk piece foregrounds DeFi as the industry with the most to lose. Smart contract code is public. Bug bounties are uneven. Once an exploit is written, on-chain execution is fast and final. There is no rollback button, no SOC analyst pulling a network cable. A capable attacker armed with a capable model is a different threat than a capable attacker alone.
Why it matters
DeFi has already lost more than $840 million to hacks in 2026, per the CoinDesk report, and the year isn't half over. That tally sits against a backdrop where most large-cap protocols are audited, multisig-protected, and have insurance backstops of varying quality. The exploits are still landing.
Drop a frontier model into that workflow and the economics shift. Reading an unfamiliar Solidity codebase, mapping its control flow, and proposing exploit candidates is exactly the kind of structured reasoning current models are getting better at. The question security researchers keep asking in private channels isn't whether Fable 5 can do this. It's whether it can do this without being asked nicely.