Been watching real adversarial input hit my detection API for six months. Here's what's actually landing.

In three linesSix months of real adversarial attacks against a prompt injection detection API (Bordair). Three dominant patterns: multi-turn setups invisible in isolation, forward-momentum exploitation, and role redefinition leveraging model helpfulness. Single-message classifiers consistently fail. Stateful defenses recommended over classifier-only approaches.

Summary generated by Claude — human-verified