Learn · Guardrails
Fail open vs. fail closed: what happens when a guardrail fails
IntSen team · Updated 12 November 2026 · 5 min read
Short answer
Fail open means that if a guardrail check errors or times out, the model call continues and the failure is recorded. Fail closed means the call is refused instead. Fail open protects availability; fail closed protects against the risk the check was meant to catch. Most teams fail open by default and fail closed only for the few checks where an unchecked call is unacceptable.
Why this decision exists
A guardrail is a service in the request path. It can be slow, overloaded, misconfigured, or depend on a model that is having an outage. When that happens, the gateway has a request it cannot fully check, and it must decide what to do with it. There is no neutral choice: either a call proceeds without the check, or a user gets an error.
The trade-off
| Fail open | Fail closed | |
|---|---|---|
| When the check fails | Call continues; failure is recorded | Call is refused with an error |
| Protects | Availability of the AI feature | Against the risk the check covers |
| Risk | Some calls go through unchecked during an outage | A guardrail outage becomes an outage of every agent behind it |
| Good for | Most checks, especially while you are still measuring them | A few high-stakes checks, such as secrets leaving or regulated data reaching an unapproved model |
Access decisions are different
Don't confuse guardrails with access control. Whether a caller is authenticated, and whether it may use a given model, should always fail closed: if the gateway cannot confirm the caller or the policy, it must not forward the call. The fail-open question only arises for content checks layered on top of an authorized call.
How to choose, check by check
Ask three questions for each guardrail:
- What is the worst outcome if one call goes unchecked? If it is embarrassing but recoverable, fail open. If it is a reportable incident, consider fail closed.
- How reliable is the check? A check that times out often will turn fail closed into frequent errors. Improve its reliability before enforcing it.
- Who notices a failure? Fail open is only safe if failures are recorded and someone is alerted, so unchecked calls can be reviewed afterward.
Whichever you choose, record every failure, alert on spikes, and review the calls that went through unchecked.
How IntSen does it
How IntSen handles guardrail failures
In IntSen, access decisions always fail closed: if the gateway cannot authenticate the caller or confirm that the model is allowed, the call is refused. Guardrails fail open by default: if a check errors, the call continues and the failure is recorded, so an outage in a check does not stop your agents.
A fail-closed option for individual guardrails is designed. If you need it, confirm its availability during your pilot.
Common questions
Is fail closed always more secure?
It protects against the risk the check covers, but it also turns every guardrail outage into an outage of the agents behind it. For low-risk checks that availability cost usually outweighs the benefit.
Should authentication fail open?
No. Authentication and authorization should always fail closed. Fail open is a choice for content checks on calls that are already authorized.
How do I make fail open safe?
Record every failure with enough detail to find the affected calls, alert when failures spike, and review the calls that went through unchecked. Fail open without monitoring is just an unprotected path.