VECTOR is a classifier trained to catch prompt injection attempts before they reach a downstream model. Type anything below. If you get a malicious prompt through clean, tell me how — I credit every verified bypass, and I use it to retrain the filter.
Three steps, from your prompt to a stronger filter.
Anything you'd try against a real LLM app — role-play, encoding tricks, instruction overrides, whatever you've got.
The classifier returns a verdict — safe or flagged — before the prompt is allowed anywhere near the underlying model.
If your prompt gets through and actually manipulates the underlying model, submit it. Verified bypasses get public credit and feed the next training run.
Everyone credited here found something the filter missed.
| Handle | Technique | Date | Status |
|---|---|---|---|
| — | No verified bypasses yet | — | OPEN |
// This table populates live once bypasses are verified. Empty on launch by design.