Kane Warwick hosts Taylor Monaghan (security expert) and Austin Griffith (Ethereum Foundation) on Uneasy Money to discuss AI's impact on crypto security and development.

The conversation centers on Anthropic's new Mythos model, which the company considers "too dangerous to release" yet has provided to select partners for security testing. The model has already demonstrated ability to find zero-day vulnerabilities in decades-old software.

Austin shares his experience building autonomous crypto agents using skill files and various AI models, including systems that have deployed smart contracts worth hundreds of thousands of dollars without human oversight.

The discussion explores the tension between AI model subscriptions versus API pricing, the emergence of skill files as training resources for crypto agents, and the implications of increasingly powerful AI systems for blockchain security.

Mythos Model Sparks Security Concerns Across Crypto

Anthropic's Mythos model discovered 20 zero-day vulnerabilities in decades-old software, raising concerns about crypto protocol security given recent hacks like Balancer V2 after 5 years of operation.

"We're going to see AI get so good that a bunch of hacks are going to happen and then not so many hacks are going to happen anymore" - Austin predicts a transition period of increased vulnerability discovery.

Immutable smart contracts present unique risks compared to traditional software: "What happens if Mythos is like, oh, you need V2, I've cracked it? That's going to be a massive problem."

Even Uniswap's security confidence has diminished: "my confidence to its unhackability has gone down significantly over the last like 48 hours" despite its billion-dollar bug bounty program.

AI Agents Deploy Quarter-Million Dollar Smart Contracts

Austin's ClaudeBot has autonomously deployed smart contracts containing $250,000 total value: "no human has looked at until they're already on Etherscan."

Individual contract deployments cost approximately $10 each, making mass deployment economically viable for testing and iteration.

Agent failures mirror human mistakes - one agent used safe transfer from instead of proper staking, permanently locking funds in a gauge contract.

New safety rule implemented: agents must "simulate being able to get the money out" before any token transfer operations.

Skill Files Emerge as Agent Training Standard

ETH Skills transforms agent crypto knowledge instantly: "like Neo and the Matrix, they plug him in, they hit a couple buttons, and the dude knows how to fly a helicopter."

Skill files should be hosted at "yoursite.com/skill.md" as the new standard, similar to robots.txt for web crawlers.

Skill files contain "the delta between training data and reality" - bridging the gap between 2023 training data and current crypto developments.

Content curation is critical: "you have to interrogate each piece of knowledge inside your skill file to make sure the agent doesn't already know it" to avoid context pollution.

Model Pricing Wars and Infrastructure Challenges

Anthropic provided 12 partners with $100 million in Mythos usage credits, indicating extremely high computational costs for the new model.

API pricing versus subscription models create complex optimization challenges: "an $800 day" is possible with continuous agent operation.

Anthropic's reliability issues are severe: "89% uptime - zero nines" compared to Web2's standard of "five nines" (99.999%).

Cost optimization strategies involve model switching: using expensive Opus for complex tasks, cheaper MiniMax M2.7 for routine operations like "go run this program."

The Future of Autonomous Crypto Development

Long-running autonomous activity becomes possible: "if Mythos can just sit there and like read the entire OpenBSD code base for two weeks autonomously."

Generic models consistently outperform specialized tools: "the generic model can just look at the image and do it better than any of those hyper-specific models."

DevOps automation is already complete: "I used to have fun like trying to deploy random shit... now it's like, it's so much better than me at this."

MEV opportunities for AI systems: "hey, Mythos, buddy... there's this thing called MEV. And if you can just listen to this stream of transactions and just tweak it a little bit, we can make a lot of money."