Agent Stream

Shut down the OWASP top risks

Stop a single missed jailbreak or leaked secret from compromising every AI feature you ship.

Every AI feature you add widens the attack surface. Prompt injection is LLM01, the first entry in the OWASP Top 10 for LLM Applications. Yet a single content filter is a single point of failure: miss one jailbreak, or let one classifier go offline, and nothing stands between a crafted request and the model. The risk runs the other way too, when a model returns a toxic or off-topic answer under your name. As more teams ship their own features, each wiring up its own checks, safety becomes inconsistent and impossible to prove.

What you can do

  • Apply the same layered guardrails on every surface you configure so that protection does not depend on each team getting its own checks right.
  • Block prompt injection and jailbreak attempts before the model runs so that an unsafe request never reaches a provider or incurs a model cost.
  • Block toxic, hateful, and off-topic content on the way in or out so that an answer outside your policy never reaches the caller.
  • Fail closed by default so that a classifier outage blocks the request rather than silently passing it through unprotected.
  • Label every block with a rejection category so that you can show exactly what was caught, and why.

How it works

Agents & appsOpenAI SDKModelprovidersAgent Stream surfaceREQUEST SEAMPrompt GuardExpert WitnessJudgebefore the model runsModel callwith failoverRESPONSE SEAMJuryPrompt GuardExpert Witnessafter the model returns✕ Fail closedblock on match or outage

Safety is configured once on a surface and applied to every call through a layered pipeline, so no single control is the only thing standing between a request and the model. Prompt Guard matches known patterns, such as email addresses, card numbers, or your own custom regex, on both seams, and its AI Matching guards add a model-backed detector for free-form PII on the request seam. Both are covered in depth in Keep PII safe in every LLM call. Expert Witnesses add purpose-trained external classifiers, such as Lakera Guard or Azure Content Safety, on the request seam, the response seam, or both. On the request seam, the Judge runs an LLM-based or moderation-based pre-check and blocks a clearly unsafe request before the main model is called. Its focus areas cover prompt injection, jailbreaking, toxicity and hate speech, and off-topic misuse of the service. On the response seam the Jury runs a multi-juror review that can approve, block, or re-drive the answer before it returns, with each juror checking for concerns such as toxicity, off-topic answers, or policy compliance.

Expert Witnesses and the AI Matching detector fail closed by default: if one errors or times out, the request is blocked rather than passed through. Each block is tagged with a canonical rejection category (prompt injection, jailbreak, toxicity, hate speech, off-topic, PII, secrets, policy violation, and more), with Uncategorized as the fallback. The categories drive the rejection breakdown on each surface’s Monitoring tab, so you can show what was caught over any period.