5 min readFeatured
Prompt Injection Grew Up
Indirect prompt injection has moved from research demos to real CVEs in Microsoft 365 Copilot, GitHub Copilot, Cursor, and Claude connectors. A short read on where the threat actually stands now.
Topic archive
AI Security
2 posts, newest first.
5 min readFeatured
Tokenization Attacks: The AI Security Risk Almost Nobody Models
Why retrieved text is not just ‘untrusted input’ but an attacker-controlled instruction channel when your model can call tools — and how to think about controls without magical safety guarantees.