Anthropic’s upcoming public release of its advanced AI model, Mythos, has prompted urgent security warnings to the decentralized finance (DeFi) community. Crypto analyst The DeFi Investor advised users on June 9 to revoke all token approvals, use only heavily audited dApps, and split funds across multiple wallets to reduce risks from AI-driven vulnerability discovery. The concern is rooted in Mythos’s reportedly exceptional ability to find severe software flaws—a capability that could dramatically accelerate attacks on unverified or poorly audited smart contracts.
The warning comes after Zcash lost over 35% of its value in a single day when a security researcher using an earlier Anthropic model, Claude Opus 4.8, uncovered a critical bug in its shielded Orchard pool. The flaw allowed infinite minting of ZEC tokens, leading prominent investor Arthur Hayes to exit his entire ZEC position. Mythos is expected to surpass Opus 4.8’s abilities, intensifying fears of a new wave of exploits.
Meanwhile, Chainalysis reported that DeFi protocols lost at least $36.7 million over six months to hacks focused on unverified smart contracts. The largest single theft hit Truebit, a protocol verifying off‑chain computations, for $26.2 million. An attacker exploited a vulnerability in an unverified contract that had been live on Ethereum since 2021. Other victims include Trusted Volumes, Aperture Finance, and Ekubo. Chainalysis highlighted that AI‑powered decompiler tools now allow malicious actors to reverse‑engineer bytecode and launch attacks at scale, making code transparency an urgent security requirement.
Anthropic has been running Mythos in a restricted “Project Glasswing” program with about 50 organizations, including Amazon, Apple, Google, and Microsoft, and plans to expand access to 150 more entities across 15 countries. However, the public version is expected to carry substantial guardrails. Despite this, the broader debate over DeFi’s viability has intensified: OpenZeppelin co‑founder Manuel Aráoz declared “all of DeFi unsafe” in May, while Aave Chan Initiative founder Mark Zeller argued that fewer than 10% of recent DeFi failures stemmed from code‑level bugs. Anthropic itself anticipates that AI will ultimately favor defenders, though “the transitional period will be fraught.”
