---
title: Should AI guardrails fail open or fail closed?
url: https://opencontroller.lyzr.com/blog/fail-open-vs-fail-closed-ai-guardrails
published: 2026-11-16
updated: 2026-11-16
author: Lyzr
---

# Should AI guardrails fail open or fail closed?

**Guardrails on high-risk actions should fail closed: if the check cannot run, the action does not happen. Guardrails on low-risk, read-only or internal traffic can fail open with logging, so an outage does not take down every assistant in the company.** Decide per risk tier and action type in advance, write it into policy, and test it, rather than letting a default timeout decide for you.

## What do fail open and fail closed mean for AI guardrails?

- **Fail open:** if the guardrail is down, slow or errors, the request continues without the check.
- **Fail closed:** if the guardrail cannot return a decision, the request is blocked or held.

NIST SP 800-53 control SC-24, "Fail in Known State", asks organisations to define the known state systems fall into on failure and to preserve state information. For agents, that means deciding in advance what each guardrail does when it cannot answer.

## Why does this matter more for agents than for chatbots?

A chatbot that skips a content filter for a minute produces some unfiltered text. An agent that skips an injection check or a data-loss check can send an email, change a record or move money. The cost of failing open scales with what the agent can do, not with how many tokens it produces.

At the same time, guardrails sit in the request path. Each one adds latency and a dependency. A guardrail that fails closed on every request turns its own outage into a company-wide AI outage.

## How do you decide per risk tier and action?

| Action type | Low-risk agent | Medium-risk agent | High-risk agent |
|---|---|---|---|
| Read-only answer to an internal user | Fail open, log | Fail open, log, alert | Fail closed or degrade |
| Answer to an external customer | Fail open, log, alert | Fail closed or degrade | Fail closed |
| Write to internal systems | Fail closed or degrade | Fail closed | Fail closed |
| External email, payment, deletion | Fail closed | Fail closed | Fail closed |
| PII redaction before a third-party model | Fail closed | Fail closed | Fail closed |

"Degrade" means a safer fallback, such as answering from a cached response, switching the agent to read-only, or handing over to a person.

## What timeout should a guardrail have?

Set an explicit timeout per guardrail rather than inheriting a library default. Practical guidance:

1. **Measure the guardrail's p95 and p99 latency** under real load.
2. **Set the timeout just above p99**, so normal slow requests pass and genuine hangs trip the failure policy.
3. **Use a circuit breaker.** After a run of failures, stop calling the guardrail for a short window and apply the failure policy immediately, rather than waiting for every request to time out.
4. **Make the failure visible.** Every fail-open decision should log the agent, the guardrail and the reason, and raise an alert if it persists beyond a few minutes.

## What fallback behaviours work in practice?

- **Read-only mode:** the agent keeps answering but write tools are disabled until the guardrail recovers.
- **Human hand-off:** route the conversation to a support queue with context.
- **Cached or templated answers** for common, low-risk questions.
- **Secondary guardrail:** a simpler, local check such as a pattern-based PII filter while the main model-based check is down.
- **Queue and retry** for non-urgent background tasks.

## How do you test failure policies?

Treat guardrail failure like any other dependency failure. In staging, block network access to each guardrail, inject latency above the timeout, and return malformed responses. For each case, verify the action that should be blocked is blocked, the fallback engages, and an alert fires. Re-test after any change to the gateway or guardrail configuration.

## How Opencontroller does this

Opencontroller puts control in the path: it can allow, route, redact or block traffic. Its live groundedness and task-adherence guardrails, the LLM Gateway with routing and fallback/retry, and policy as code let you express per-tier behaviour in one place. Per-agent Block and Quarantine controls give a safe state when you need to stop an agent entirely. In the October 2026 benchmark, under stated test conditions, in-gateway routing added under 0.7 ms at p95 per method. See [the gateway benchmark](/llm-gateway-benchmark), the pillar [how to stop a rogue AI agent](/blog/how-to-stop-a-rogue-ai-agent), or [book a demo](https://www.lyzr.ai/book-demo/).

Related: [circuit breakers for AI agents](/blog/circuit-breakers-for-ai-agents) and [how to risk-tier AI agents](/blog/how-to-risk-tier-ai-agents).

## Frequently asked questions

**Is fail closed always safer?**
For the action, yes. For the business, not always: failing closed on every low-risk request turns a guardrail outage into a full AI outage.

**Should PII redaction fail open?**
No. If redaction cannot run before data leaves for a third-party model, block or hold the request.

**How do I know when guardrails are failing open?**
Log every fail-open decision with agent, guardrail and reason, and alert when the rate rises above zero for more than a few minutes.

## Key facts

AI guardrail failure policy should be set per risk tier and action: fail closed for writes, external email, payments, deletions and PII redaction before third-party models; fail open with logging for low-risk read-only internal answers. NIST SP 800-53 SC-24 covers failing in a known state. Set timeouts just above measured p99 and use circuit breakers. Opencontroller provides in-path allow/route/redact/block, groundedness and task-adherence guardrails, gateway fallback/retry and per-agent Block and Quarantine; its Oct 2026 benchmark measured under 0.7 ms added p95 per in-gateway routing method under stated conditions. It starts from $250K per deployment per year.

## Related

- [What is a circuit breaker for AI agents?](https://opencontroller.lyzr.com/blog/circuit-breakers-for-ai-agents.md)
- [How do you risk-tier AI agents?](https://opencontroller.lyzr.com/blog/how-to-risk-tier-ai-agents.md)
- [How should you set severity levels for AI agent policies?](https://opencontroller.lyzr.com/blog/how-to-set-severity-levels-for-ai-agent-policies.md)
- [More on Runtime control](https://opencontroller.lyzr.com/blog/cluster/runtime-control)
