I've been researching prompt injection and collecting real attack data. 1,400+ attempts so far. The finding that surprised me most: the attacks that actually bypass detection aren't technical at all. No "ignore previous instructions."…
Posted here yesterday about the v1 cross-modal dataset. One of you suggested adding GCG adversarial suffixes and multi-turn attack coverage. That feedback turned into v2 and v3 being built and shipped within 24 hours. The dataset has gone from 47…
I've been researching what happens when you split a prompt injection across multiple input modalities instead of putting it all in one text field. The short answer: per-channel detection breaks completely. The idea is simple. Instead of sendi…