How does a CSRF token actually stop a forged request?

The server issues a unique, hard-to-guess token per session/request and requires it to be present (matching) on every state-changing request. A malicious third-party page cannot read this token (due to Same-Origin Policy preventing it from reading responses/cookies belonging to a different origin), so any forged request it sends is missing the token or has an invalid one, and Spring Security's CsrfFilter rejects it with HTTP 403.