Lat builds deliberately vulnerable applications and teaches you to attack them: over-permissioned identities, functions exposed without controls, open storage buckets, leaked credentials, privilege escalation. Then he shows the fix. It covers the three big clouds, which is its commercial virtue and its technical limit.
What I took from it
The rule fits in a line: a component's permissions are the size of the damage it can do. It does not matter how well the code is written; if the identity it runs under can read every bucket in the account, one flaw reaches all of it. Since reading it, when I review a deployment the first thing I look at is not the code but what the executing identity is allowed to do, and the second is whether anyone has ever checked.
The second thing is more concrete and more common than it sounds: storage buckets holding company documents. Any project that puts AI over a company's own documents ends up storing contracts, invoices and email in a cloud bucket, and that bucket is precisely the object this book spends a chapter teaching you to exploit. It is the least glamorous piece of the project and the one most often misconfigured.
The third: audit logs only help if someone reads them. The book spends pages tracing an attack through the logs, and the inverse conclusion is the useful one — if nobody owns the job of reading them, they exist for the autopsy, not for the defence.
What I left out
The labs themselves, which are long and cost real money in a cloud account. And it has to be said: covering three providers means each gets a third of the depth, and console screens change faster than a book gets printed. The part that does not expire is the reasoning about identities and blast radius.