Question 1

What does Redline AI do?

Accepted Answer

Redline runs real attacks against your AI agent and everything it connects to — its tools, its MCP servers, its skills — before it goes to production, then protects every live session after it does. It tests the assembled agent rather than the underlying model, because the vulnerabilities in a deployed agent are overwhelmingly in the assembly.

Question 2

How accurate is Redline's prompt-injection detection?

Accepted Answer

On BIPIA, the indirect-injection benchmark, Redline Guard v5 reached 0.9484 recall at a 1.4% false-positive rate, against PIGuard's 0.8320 at 10% and ProtectAI v2's 0.1184. Measured 5 September 2026 on identical rows in a single run, scored on the quarter of BIPIA held back from training by hash. It loses to ProtectAI v2 on three of the seven external sets — Mindgard, NeuralChemy and Gandalf — and the full table is published on /guard.

Question 3

Does my agent have to be deployed to Redline?

Accepted Answer

No. The SDK connects outbound and pulls its work, so the agent runs where it already runs, with your own dependencies and your own model keys. There is no endpoint to expose and no proxy in front of your traffic; it works from a laptop behind NAT.

Question 4

What is prompt injection?

Accepted Answer

Prompt injection is when text a language model was supposed to read as data gets acted on as an instruction instead. Instructions and content arrive in the same context window as one stream of tokens with nothing marking which is which. Indirect injection — the payload arriving inside a document or tool result the agent read — is the more dangerous form, because the attacker is not the user.