Mythos, Anthropic’s new AI model that has sparked fear and confusion in traditional technology and finance, is also driving a massive shift in how the crypto industry thinks about security.
For years, decentralized finance has focused its defenses on smart contracts. Code is audited, vulnerabilities are cataloged, and many common exploits are well understood. But Mythos, a model designed to identify and sequence weaknesses in systems, draws attention beyond the code and toward the infrastructure that supports it.
“The biggest risks are in infrastructure,” said Paul Vijender, head of security at Gauntlet, a risk management company. “When I think about AI-based threats, I am less concerned about smart contract exploits and more focused on AI-assisted attacks against the human and infrastructure layers.”
This includes key management systems, signing services, bridges, Oracle networks, and the cryptographic layers that connect them. These components are less visible than smart contracts and often fall outside the scope of traditional auditing.
In fact, this month, web infrastructure provider Vercel, used by many crypto companies, disclosed a security flaw that could have exposed customers’ API keys, prompting crypto projects to alternate credentials and review their code. Vercel traced the intrusion to a compromised Google Workspace login via the third-party AI tool Context.ai, used by an employee.
Mythos belongs to a new class of AI systems designed to simulate adversaries. Instead of looking for known bugs, it explores how protocols interact, testing how small weaknesses can be combined into real-world exploits. This approach has attracted attention beyond cryptography. Banks like JP Morgan are increasingly treating AI cyber risk as systemic and exploring tools like Mythos for stress testing. Earlier this month, Coinbase and Binance both reportedly contacted Anthropic to test Mythos.
Early findings from models like Mythos identified weaknesses in the behind-the-scenes systems that keep crypto platforms secure, including the technology that protects keys and manages communication between systems.
“I think there are two areas where AI models are particularly valuable,” Vijender said. “First, multi-step exploitation chains that historically are only discovered after money has been lost. Second, vulnerabilities in the infrastructure layer that traditional audits never touch.”
This change is important in a system built on composability, where DeFi protocols can connect and build on each other.
DeFi protocols are designed to interconnect. They share liquidity, rely on common oracles, and interact across layers of integrations that are difficult to map in their entirety. This interconnectivity has driven growth, but it also creates pathways for risk to spread, as shown by recent bridging exploits like the Hyperbridge attack, in which an attacker created $1 billion worth of Polkadot tokens bridged to Ethereum by exploiting a flaw in the way cross-chain messages were verified.
“Composability is what makes DeFi capital efficient and innovative,” Vijender said. “But it also means that a minor vulnerability in one protocol can become a critical exploitation vector with the potential for contagion to the entire ecosystem.”
Without AI, these dependencies are difficult to trace. Thanks to AI, they can be mapped and exploited on a large scale. The result is a shift from isolated exploits to systemic failures that ripple across all protocols.
Evolution of AI attacks
Still, some industry leaders view Mythos as an acceleration rather than a turning point.
At Aave Labs, founder Stani Kulechov said AI reflects the dynamics already at play in DeFi’s contentious environment.
“Web3 is no stranger to well-funded and motivated adversaries,” he told CoinDesk. “AI models represent an evolution in the tools used to perform exploits.”
From this perspective, DeFi is already designed for machine-speed attacks. Smart contracts execute automatically and defenses such as liquidation mechanisms and risk settings operate without human intervention.
“DeFi works at computing speed, so AI does not introduce new dynamics,” Kulechov said. “This intensifies an environment that has always required constant vigilance.”
Despite this, Aave sees AI surfacing new categories of vulnerabilities, including issues that human auditors might have previously deprioritized.
“The Mythos paper shows that AI can discover old bugs that were previously low priority,” he said.
This scale is still important in a system where even smaller vulnerabilities can undermine trust or be combined into larger exploits.
If attackers can move faster, the question becomes whether defenses can keep up.
For both Gauntlet and Aave, the answer lies in changing the security model itself. Pre-deployment audits and post monitoring were designed for human-paced threats. The AI compresses this timeline.
“To defend against offensive AI, we will need to take an AI-centric approach where speed and continuous adaptation are key,” said Gauntlet’s Vijender. This includes continuous auditing, real-time simulation, and systems built with the assumption that violations will occur.
A “better way”
Aave has already integrated AI into its workflows, using it for simulations and code reviews alongside human auditors. “We’re taking an AI-first approach where it adds clear value,” said Aave Labs’ Kulechov. “But it complements, rather than replaces, human-led auditing.”
In this sense, AI equips both attackers and defenders.
For manufacturers, the long-term effect could be less disruption than divergence.
“We haven’t tested Mythos yet, but we’re genuinely interested in what it and similar tools can do for protocol security,” said Hayden Adams, founder and CEO of Uniswap Labs. “AI provides manufacturers with better ways to test and harden systems.”
Over time, Adams expects the gap between secure and insecure protocols to widen.
“Projects that prioritize security will have a greater ability to test and harden systems before launch,” he said. “Projects that fail to do so will be most at risk.”
Perhaps this is the real change. Security is no longer about eliminating vulnerabilities. It’s about continually adapting to a system in which these vulnerabilities are constantly being rediscovered and recombined.
Learn more: Move beyond Bitcoin and quantum risks. Anthropic’s AI Mythos Could Have Major Implications for DeFi
