Why Mythos is so scary, and Anthropic cuts off OpenClaw

Summary

• Anthropic's new frontier AI model "Mythos" is considered so dangerous it is withheld from public release, given only to ~40 select organizations (e.g., AWS, Apple, Nvidia) for defensive cybersecurity work under "Project Glasswing."
• Mythos demonstrated the ability to find ~20 zero-day vulnerabilities in decades-old software (e.g., OpenBSD), signaling a paradigm shift in offensive and defensive security capabilities.
• The model's power suggests a future where expensive, generic frontier models can "one-shot" complex tasks (like hacking a smart contract) without the need for intricate agent harnesses or loops, potentially collapsing current development workflows.
• For crypto, this raises existential security concerns, highlighted by recent hacks (e.g., Balancer) on old, immutable code. While some argue web3's bug bounties and audits may make it more resilient, the threat to critical infrastructure (like validators) and MEV is significant.
• A key defensive innovation is the "skill file" (e.g., ET Skills), a text document that dramatically improves an AI agent's domain-specific knowledge by updating it on post-training developments, acting like instant upskilling.
• There is a constant tension and cost optimization challenge between using expensive, high-intelligence frontier models (like Opus) and cheaper, less capable models (like Haiku or local models), with developers "oscillating" between them based on task complexity.
• Anthropic is actively cutting off API "loopholes" like OpenClaw from using subsidized subscription plans, forcing users onto more expensive, metered API pricing, which changes the economics of running AI agents.
• The rapid advancement of models creates a moving target for builders; the advice is to "build for the model six months from now," as specialized tools and harnesses may become obsolete when a smarter generic model arrives.