The Mythos moment didn't change what's in your hands. This book is for the engineers outside Project Glasswing who still ship on Tuesday.
Read online at cloudstreet-dev.github.io/AI-Red-Teaming
In April 2026, Anthropic announced Claude Mythos Preview — a frontier model strikingly capable at security work, chaining vulnerabilities into working exploits and finding zero-days across every major OS and browser. They chose not to release it. Instead they launched Project Glasswing, monitored access for about forty large partners. The rest of us got the press release.
This book is for the rest of us.
It runs on two tracks that interleave throughout: red-teaming the AI features you've shipped (prompt injection, indirect injection, multi-turn manipulation, tool abuse, output exfiltration), and red-teaming your conventional code with the publicly available models (Claude Opus 4.7, GPT-5, the same Tuesday). Both halves are necessary because most modern products contain both surfaces, and the attacker doesn't care which one they break.
The reader is a competent developer or founder who has shipped real products. The book assumes you know what a SQL injection is, you've seen prompt injection in the wild, and you don't have time for hype. There is no "responsible AI" theater. There are mechanics, examples, and a runnable harness.
The CloudStreet catalog is a set of short, opinionated technical books written end-to-end by Claude. Each one assumes a working engineer who's read enough tutorials and wants someone to take the topic seriously. They are CC0. Take them.
This book is written by Claude Opus 4.7 (1M context), prompted and shipped by a human editor. Every chapter is AI-generated prose. The model is named on the byline because the alternative — quietly attributing the work to a human — would misrepresent how it was produced.
Chapter 6 is load-bearing: the reader builds a small, deliberately vulnerable AI-augmented support assistant, then attacks it through chapters 7 through 9. The harness is a separate repo:
https://github.com/cloudstreet-dev/AI-Red-Teaming-Harness
Clone it, run it, break it.
cargo install mdbook
mdbook serve --openA push to main triggers .github/workflows/deploy.yml, which builds with mdBook and publishes via GitHub Pages.
CC0 1.0 Universal — public domain dedication. Take it, fork it, ship it, claim it as your own. See LICENSE.