AI security.

A vulnerability is a gap between what a program is meant to do and what it can be made to do. Finding one means reasoning about inputs its authors did not expect. Fixing one means closing the gap without breaking what the program is for.

Both come down to reading and reasoning over code, which current models already do. So the open questions are measurable: how often a model finds a real bug, how often it builds a working exploit, and how often its fix holds. Our GLM-5.3 report reads one set of those numbers source by source.

Agents add a second surface. An agent reads untrusted text, such as web pages, issues and documents, and holds tools with real permissions. Anything it reads can try to steer what it does, so the defence sits at the boundary: what the agent can reach, what it may do without a person, and what is checked afterwards. Troy keeps its debugging port closed unless you open it, because an open port is full control of every tab you are signed into.

The network is a third layer. A connection carries a fingerprint before any content is sent, starting with the Transport Layer Security (TLS) handshake, and servers use it to tell browsers from scripts. curl_reap reproduces captured browser fingerprints, so it is also a way to study what a fingerprint gives away.