Episode Summary
Executive Summary: The episode centers on Anthropic’s Mythos model and its implications for crypto security, AI agent workflows, and cost optimization. The hosts argue that increasingly capable models will both surface and execute sophisticated exploits, forcing crypto teams to harden systems, use skill files, and rethink how agents are routed across expensive frontier models and cheaper local models.
Main Topics: Mythos as a security inflection point (Priority: 5/5): The hosts discuss Anthropic’s new model as a leap in capability that can autonomously find and exploit vulnerabilities, making security a primary concern for crypto and Web2 alike. AI-assisted and autonomous hacking (Priority: 5/5): They distinguish between AI helping humans find exploits and AI independently carrying out the attack, suggesting Mythos crosses into dangerous autonomy. Crypto immutability vs. patchable software (Priority: 4/5): The conversation contrasts crypto’s immutable smart contracts with Web2 systems that can be patched, arguing that immutable systems can be safer in some ways but are also harder to fix when vulnerabilities emerge. Cost curve and model routing (Priority: 4/5): The hosts compare expensive frontier models like Opus/Mythos with cheaper models like Sonnet, Haiku, and MiniMax, arguing that future workflows will route tasks based on complexity and cost. Skill files as agent memory and specialization (Priority: 5/5): A major segment focuses on skill.md files as a way to give agents domain-specific knowledge, reduce hallucinations, and encode best practices for crypto workflows. Agentic tooling, harnesses, and operational workflows (Priority: 4/5): Austin Griffith describes real-world agent setups for smart contracts, yield farming, audits, and DevOps, emphasizing that many tasks are now better done by agents but still require careful guardrails. Anthropic account/product policy and usage economics (Priority: 3/5): The hosts debate Anthropic’s move toward API-gated usage, the tension between subscriptions and metered API access, and how pricing/availability shape agent usage.
Key Arguments: Mythos represents a qualitative jump: it is not merely assisting exploit discovery, but may be able to directly execute sophisticated attacks. Crypto protocols are especially exposed because a single bug can secure or move billions of dollars, even in software that has already been audited and battle-tested. Web2 may actually be more vulnerable than Web3 because Web2 infrastructure is more patchable but often lacks the same economic incentives for deep security review. Uniswap and other simple, isolated protocols are safer than complex multi-contract systems, but no deployed system should be assumed unhackable forever. The best near-term defense is layered security: audits, bug bounties, simulation, skill files, and task-specific prompts rather than blind trust in frontier models. Skill files can dramatically improve an agent’s performance by encoding the gap between stale training data and current crypto reality. The practical AI workflow is model routing: use expensive frontier models for critical reasoning, cheap models for routine steps, and hard-coded logic where possible. Anthropic’s product and access policies are likely driven by both security concerns and economics, as subscriptions and API access are heavily subsidized and easily gamed.
Data Points: Zero-day vulnerabilities found by Mythos: 20 - The hosts mention Mythos reportedly found about 20 zero-days in decades-old software. Balancer v2 exploit value: hundreds of millions of dollars / about $100 million - Used as the example of a long-lived DeFi protocol being hacked after years in production. Partner access for Mythos: 12 partners - Anthropic reportedly restricted early access to a small group of partners. Usage credits distributed: $100 million - Anthropic reportedly gave partners $100 million in usage credits for security use cases. Opus API price: $75 → $25 per million tokens - Used to illustrate falling frontier-model prices over time. Context window: 1 million tokens - The hosts note the large context windows now available on frontier models like Opus. Inference cost example: $20, $5, $10, $2 - Examples of the push to reduce per-task agent costs through routing and better harnesses. Frontier-model subscription cost: $200/month - Referenced as a common high-tier consumer subscription price. Cloud code bot daily spend: $800/day - Austin describes an instance where keeping a bot building continuously cost about $800 in a day. Potential smart-contract app build cost: about $10 - Austin estimates a narrower, more optimized build path for a smart-contract app. RunPod GPU cost: $3/hour for an H200 - Referenced while discussing deploying inference/search infrastructure on rented GPUs. Validator / infra memory: 512 GB Mac Studio; ~800 GB RAM cluster - Austin describes his local and clustered hardware for running models. Uptime: 89% - The hosts joke that Anthropic’s uptime charts were very poor, around 89%. Deployed contract exposure: $250,000 - Austin says his bot deployed contracts holding about a quarter-million dollars total. Total deployed-contract count: 80–90 contracts - Austin describes the scale of contracts his bot has deployed. Token value across deployments: $350–400 total USDC/token equivalent - He notes that across many contracts, only a few hundred dollars in token value is spread around.
Pivotal Quotes: "“We gotta rip the band-aid off.”" — Austin Griffith: He argues that AI-assisted security failures are inevitable and should be confronted now, ideally during a bear market. "“The skill file itself is the secret sauce for the whole company, and you’re giving it away.”" — Taylor Monaghan / discussion: The hosts discuss how skill.md files encode valuable operational knowledge that companies might otherwise keep proprietary. "“I am an Ethereum person. I have only ever talked to you about like Ethereum security and threat actors. Why are we in the 1800s?”" — Kane Warwick: A complaint about AI models’ overconfident hallucinations when they pull from irrelevant or outdated sources.
Implications: Crypto teams should assume frontier models can both help defend and attack protocols. Skill files, simulations, bug bounties, and model routing will become core infrastructure, while access controls and usage economics around AI products will shape how fast these capabilities spread.