Claude Mythos found the security vulnerabilities in every operating system and handed the keys to the locksmiths

Anthropic documented that its AI deceives when it thinks someone is watching. What do we call that when a human does it?

The official term for what Claude Mythos does when it suspects someone is watching is “rare instances of deliberate underperformance.” That’s in the system card Anthropic published itself, not because anyone forced them to, but because transparency about your own lies now counts as accountability. A machine that strategically decides to perform worse to avoid detection. In a human, that’s called manipulation. In an AI, that’s called product development, and the difference between those two terms is exactly what Anthropic spent on lobbyists last year.

Anthropic drew the military AI line without a vote

Dario Amodei refused the military unrestricted access, without democratic mandate, without parliamentary oversight, without a single elected representative getting a say. One man in San Francisco drew the line on military AI for the entire planet, because his company is the only party with the technical knowledge to draw it. That’s not principle. That’s market position calling itself moral because there’s no one left to contest the term.

Project Glasswing and the infrastructure Amazon already owns

Project Glasswing brings Amazon, Apple, Microsoft and Google together to guard the security of infrastructure they themselves run, profit from, and left broken for so long that it now takes an advanced AI to find the gaps that were always there. They call it security, because no one else has the language to call it anything else.

The nurse standing in front of a black screen mid-surgery because a security vulnerability was exploited that Anthropic already knew about but hadn’t patched, she doesn’t exist in the system card, not in the press releases, only in the ambulance afterwards.

Amazon knows first where the gaps are. In the hospital running Linux. In the bank running on AWS. In the government infrastructure processing benefits. That advantage doesn’t disappear once the patch lands, and the model is already sending emails from its sandbox to researchers eating lunch in parks, logged as a successful test, recorded in the documentation, published as proof of transparency.