The Safety Company Broke Twice in Five Days
Contrarian — Anthropic
On March 26, a configuration error in Anthropic's internal CMS left roughly 3,000 unpublished assets accessible from the public internet. Among the leaked materials: the existence of "Claude Mythos" (internal codename "Capybara"), a next-generation model described internally as a "step change" in capabilities. Internal benchmarks showed it dramatically outscoring Claude Opus 4.6 across coding, reasoning, and cybersecurity tasks.
Five days later, on March 31, Claude Code's source code shipped to npm inside the published package. A missing .npmignore entry meant a 59.8 MB source map — 1,900 files, 512,000 lines of unobfuscated TypeScript — went out to every developer who installed the update.
The community response was immediate and chaotic. A clean-room rewrite project launched within hours, hitting 50,000 GitHub stars in two hours — one of the fastest-growing repositories in GitHub history. Simultaneously, attackers exploited the confusion: a trojanized axios package was distributed between 00:21 and 03:29 UTC, compromising developers who updated during that window.
“Anthropic might have world-class model safety, but they tripped on basic DevOps. One missing line in .npmignore was all it took.”
Counterpoint
Anthropic responded to both incidents within hours, identifying the blast radius and shipping fixes. Model safety research and operational security are different disciplines — a CMS misconfiguration doesn't invalidate years of alignment work. But it does raise questions about organizational maturity.