Anthropic Accidentally Told the World Its AI Can Break the Internet

1 min read Original article ↗

3,000 files. No password. A model that “far outpaces defenders.” And an IPO in six months.

Delanoe Pirard

Press enter or click to view image in full size

Picture a safe manufacturer who publishes the blueprints for his “unpickable” lock on a Google Doc set to “anyone with the link.”

TL;DR

  1. A misconfigured CMS exposed ~3,000 internal Anthropic assets, revealing the existence of Claude Mythos (Capybara tier), described as a “step change” over Opus 4.6. (Fortune, 03/26)
  2. Mythos is “far ahead of any other AI model in cyber capabilities.” Cybersecurity stocks dropped 4.5% to 9% in a single session. (CNBC, 03/27)
  3. The RSP v3.0, which removed Anthropic’s pause commitment, had been published 30 days before the leak. (Anthropic RSP v3.0)
  4. ASL-4, the safety level that should cover a model of this power, has not been defined yet.
  5. Anthropic is targeting an IPO at $60B+ in Q4 2026. The timing of the leak, accidental or not, amounts to free worldwide marketing. (The Tech Portal, 03/27)

I. The Locksmith Who Left His Door Open

Cybersecurity stocks lost between 4.5% and 9% on a Friday morning. CrowdStrike, Palo Alto, Zscaler, SentinelOne, Okta, Tenable. All down. Because someone at Anthropic forgot to check a box.