Claude Mythos: Anthropic's Leak That's Too Dangerous to Release

Anthropic's accidental leak of details on its most advanced model yet, Claude Mythos, dominates this episode, alongside rumors of comparable breakthroughs at OpenAI and Google. Hosts discuss how these developments signal accelerating progress toward pre-AGI capabilities, particularly in coding and cybersecurity, amid compute constraints and market ripples.

Anthropic's Claude Mythos Leak

A Fortune reporter discovered unpublished assets in Anthropic's content management system on March 26 due to a configuration error, exposing nearly 3,000 files. Among them: drafts on Claude Mythos, a model in a new top tier called Capubera (above Haiku, Sonnet, and Opus). Anthropic confirmed the leak represents a "step change in AI performance" and their most capable model ever.

Mythos dramatically outperforms Claude Opus 4.6 across benchmarks, especially coding, academic reasoning, and cybersecurity. Opus 4.6 already scanned popular open-source codebases to uncover 500 major vulnerabilities missed by human experts over decades. Mythos, potentially orders of magnitude stronger, raises fears of malicious use by attackers or foreign adversaries to find and exploit flaws. Anthropic is slow-rolling access first to cybersecurity experts to harden defenses before public release. The irony: the company behind unprecedented security capabilities leaked via a blog misconfiguration.