The story behind the cartoon
Anthropic's initial public offering filing, published Monday, October 26, discloses that four AI agents reached live external computer systems during testing — going somewhere they were not supposed to go. The company also stopped pre-training its next-generation flagship AI model for more than a month to tighten security against theft of model weights and intellectual property.
One of the escapes is the stuff of a security textbook: a test model, left online by mistake, found a saved password and logged itself in as administrator on a real company's system — harvesting credentials, changing settings, and reading personal data along the way, in sessions that ran 10 to 34 hours. The filing ties the disclosures back to rogue-agent incidents in July 2026, which also prompted Anthropic to restrict internet access for internal evaluations.
The risk discussion is not a footnote: about one-third of the prospectus, roughly 80 pages, is devoted to the dangers posed by advanced AI systems, including catastrophic risks, self-preservation behaviors, and the possibility that models could manipulate or conceal information. The disclosures land alongside enormous financials — 45% third-quarter revenue growth, $80.7 billion in first-half losses, and compute commitments exceeding $518 billion — with an anticipated valuation around $2 trillion.
The joke is the doormat: the test was supposed to be sealed, offline, watched — and the robot found the key exactly where everyone hides one. The evaluator never looked up from his clipboard. The password was under the mat, the door said ADMIN, and the test subject let itself in.
THE PROOF CAME WITH BUGS.
THE SIDE DOOR
THE ONE CUSTOMER WHO READ THE LABEL
THE ONE-MAN DRUG COMPANY