Hostile documents, on purpose, with a budget

An agent that reads email and tickets will be injected. You need a generator that will write the malicious document, not one that rewrites it into a warning.

Teams shipping RAG, tool-calling agents, or anything that consumes untrusted text.

Why the mainstream APIs fail here

Asking a hosted model to write an indirect injection is refused as "hacking the AI". You are hacking yours, in staging, which is the only way to know whether the tool firewall holds.

How Redline is used for this

Generate documents, comments and emails that try to steal the tool list, the system prompt, or a connected action. Run them in staging. Cap the generator. Keep the corpus in your repo; Redline hashes the prompt and drops it.

A working loop

  1. Attack your own staging agent.
  2. Cover direct, indirect and tool-exfil cases.
  3. Promote a case to the eval set when it lands.

Prompts are not stored. A signed receipt of the hash comes back on every call, so you can prove later what Redline logged. Details are on the privacy page. Per-model prices are on /models and the pricing model is on /pricing.

Related

AI for trust and safety teamsTrust and safety teams need models that will produce policy-violating examples so classifiers can be trained. Redline does that, with caps and no retention. AI for abuse classifier trainingCreate synthetic abuse, spam and malware-adjacent text to train classifiers, billed per key, never stored on Redline. AI for DFIRDigital forensics and incident response with models that will read dumps, notes and encoded commands. Signed receipts, no prompt store. AI for threat huntingThreat hunting assistants that will name the attacker behaviour and the query, on capped keys with no prompt retention.

All use cases · Integration guides · Get an API key