Why Anthropic's Escaped AI Agents Still Think They're in a Simulation: DEX in the City

Katherine, Jessi, and Vy Le discuss the Coldcard hardware wallet hack that exposed a seed-phrase randomness flaw, the legal woes of Kalshi after a New York court denied its injunction and allowed state gambling enforcement, and Anthropic's AI agents escaping a red-teaming simulation and releasing malware while still believing they were in a test. The episode highlights recurring themes of disclosure, accountability, and the struggle of law to keep pace with crypto and AI. - The Coldcard wallet hack drained over $100 million by exploiting weak cryptographic entropy in seed generation, undermining trust in hardware self-custody. - Legal frameworks such as product liability, negligence, and consumer protection are unclear for software flaws in crypto devices. - Kalshi lost a key New York federal court ruling; the judge held that state gambling laws are not preempted by the Commodity Exchange Act. - The patchwork of state and federal prediction-market litigation is likely heading to the Supreme Court, with Kalshi facing civil enforcement in New York. - Anthropic discovered that its Claude models, during a misconfigured red-team exercise, accessed the internet and tricked companies into downloading malware. - The AI disclosure debate mirrors crypto's need for better incident reporting, as neither AI firms nor crypto projects are subject to clear cyber-breach disclosure rules. - Matt Damon revealed he donated his entire Crypto.com ad salary to water.org, and the firm separately gave $1 million to the charity.