DeFi security has spent the better part of a decade fixated on one thing. Smart contract code. Audit it, catalog the bugs, ship the fix. That defense model is now being walked past by the same AI systems the rest of the industry is still trying to figure out.
Anthropic’s Mythos model, the one the company has called too dangerous for public release, is exposing where DeFi was actually weak the whole time. Not in the contracts. In the infrastructure layer underneath, the part nobody pays much attention to. Key management systems. Cross-chain bridges. Oracle networks. The plumbing. Mythos chains small weaknesses across these pieces into systemic, cascading failures, the kind of attack that historically only gets discovered after the money is already gone. April has been a brutal month for that exact pattern. DeFi protocols lost over $578 million in 18 days to one hacking unit, and the bridge-and-oracle layer kept showing up in the postmortems.
Mythos does not look for known bugs. It looks for combinations.
Older AI security tools scan code for patterns. They flag the same SQL injections and reentrancy bugs that human auditors already know about. Mythos belongs to a different class. It simulates an adversary. It explores how protocols actually interact, then tests whether small weaknesses in two or three different places can be combined into a real-world exploit.
In Anthropic’s own testing, the model identified vulnerabilities in widely deployed cryptographic libraries, including bugs that had sat in code for over twenty years. One example was a 27-year-old bug in OpenBSD, an operating system whose entire reputation is built on security. It also wrote a working browser exploit that chained four separate vulnerabilities together. Over 99% of what the model has found is still unpatched, which is why the public reporting is limited.
That capability profile is exactly the gap DeFi has been losing money in.
Banks are testing it. So are Coinbase and Binance.
JP Morgan is already treating AI-driven cyber risk as systemic and exploring tools like Mythos for stress testing. According to early reporting, both Coinbase and Binance have approached Anthropic in recent weeks to test Mythos against their infrastructure. That is the part the rest of the industry should pay attention to. The largest centralized players are not waiting for the technology to mature. They are running it against their own systems before someone else does. Coingo previously covered Coinbase’s app store for AI agents, which signaled how seriously they were treating AI-native infrastructure. Mythos testing is the same posture, applied to security.
Access to Mythos right now is limited to a few dozen owners of critical infrastructure. Defensive work is being prioritized before any broader release. The strategic logic is straightforward. Find the bugs and fix them, before the offensive version of this capability reaches the public.
The shift from contract bugs to infrastructure risk
Here is the practical contrast that DeFi security teams are now staring at.
| Old DeFi Defense Model | What Mythos Targets Instead | Examples |
|---|---|---|
| Audit smart contract code | Multi-step exploit chains across protocols | Bridge + oracle + lending stack |
| Catalog known bug classes | Subtle weaknesses in key management infrastructure | Signing systems, HSMs |
| One-time security review | Continuous AI-driven adversarial testing | Stress tests, simulations |
| Single-protocol focus | Cross-protocol interaction risk | Bridges, oracles, composability |
The problem with the old model was never that audits were wrong. They were just incomplete. A smart contract can be flawless on its own and still get drained because the oracle feeding it price data has a vulnerability, or the bridge moving its tokens has a different one, or the key management system sitting behind both has a third. Mythos is built to find that exact category.
AI is going to arm both sides
DeFi leaders are converging on the same forecast. AI will arm attackers and defenders simultaneously. The gap between protocols that adopt continuous, AI-driven auditing and those that stick with quarterly human reviews will widen fast. The industry has already been getting a preview of what one side of that fight looks like. The North Korean IT worker operations inside DeFi have been running multi-step infiltration attacks for years, and the $285M Drift hack is the kind of cascade that an AI like Mythos is purpose-built to find before it happens. Or to find first, depending on which side gets there.
Alchemy CEO Nikil Viswanathan has been making a related argument from the other end. The financial system was designed for humans, but the next wave of activity will be driven by AI agents that operate natively in crypto, where the borderless and continuous nature of the rails matches how agents actually work. If he is right, the systems being stress-tested by Mythos right now are the same systems those agents will be transacting on in two years.
That is the part the public has not absorbed yet. The defensive testing happening this month is not just protecting current users. It is building the security baseline for whatever the next layer of users turns out to be.
Whether the rest of DeFi catches up before that becomes the new normal is the open question.
