Anthropic will reportedly release a public version of its Mythos AI model, and crypto analyst The DeFi Investor is urging decentralized finance users to take action before that happens.
The concern is based on how good Mythos is at finding software vulnerabilities, and a version of it that becomes widely accessible could accelerate the speed at which attackers discover and exploit weaknesses in DeFi protocols.
What the DeFi community needs to do
In a June 9 post on
For those unfamiliar, token approvals are permissions that users give to smart contracts, allowing the contracts to issue tokens on their behalf. They tend to accumulate silently over time and provide a permanent attack surface if an approved contract is later found to be vulnerable.
“The scary thing about Mythos is that it is insanely good at finding serious vulnerabilities,” wrote The DeFi Investor. “Claude Opus 4.8 also recently identified a critical bug for Zcash, and Mythos is said to be even better than Opus 4.8.”
They added that DeFi will undergo a massive stress test in the coming months, and indeed, the Zcash vulnerability they mentioned provided a concrete illustration of this.
The privacy coin lost more than 35% of its value in one day after a security researcher using AI discovered a bug in the gated Orchard pool that allowed bad actors to mint new coins endlessly. $ZEC tokens. It saw major crypto investor Arthur Hayes exit his entire company $ZEC position, as uncertainty grew about whether someone had already exploited the flaw.
Mythos has been restricted to around fifty organizations, including Amazon, Apple, Google and Microsoft, since April through an anthropic initiative known as Project Glasswing, in an attempt to use the model’s capabilities for defensive purposes. According to Bloomberg, Anthropic plans to expand that circle with another 150 organizations in 15 countries.
However, multiple sources, including TFTC and journalist Alex Heath, have claimed that the public version of Mythos will have “substantial guardrails” and will not be as permissive as what Project Glasswing partners have access to.
A debate that DeFi was already having
DeFi Investor’s security tips come at a time when a conversation is emerging around the viability of decentralized finance.
In late May, OpenZeppelin co-founder Manuel Aráoz declared “very DeFi unsafe” and said he had advised people to exit positions in major protocols including Aave, MakerDAO and Compound. His reason for doing so was that AI has tilted the balance of security so far toward attackers that no protocol can currently be trusted to safely hold users’ funds.
And indeed, many crypto projects have been hit in recent months, including attacks on KelpDAO and the Drift Protocol in April, which resulted in a combined loss of more than $570 million. More recently, hackers reportedly extracted at least $30 million worth of the Humanity Protocol’s H token from 17 wallets.
However, according to Mark Zeller, founder of the Aave Chan Initiative, concerns about AI are overblown: Less than 10% of DeFi security flaws in the past year were caused by code-level vulnerabilities.
Anthropic’s own position, according to Bloomberg, is that AI will favor defenders in the long term, but “the transition period will be tough.”