# What is Fail Open and Fail Closed?

Also called fail-open, fail-closed, fail safe, fail secure.

Fail open and fail closed describe the two ways a system can behave when a check or dependency is unavailable. Failing open allows the operation to proceed, favoring availability. Failing closed blocks the operation, favoring safety or security. The correct choice depends on what the failing component protects, and picking the wrong default is a recurring source of both outages and breaches.

The terms come from electrical and physical security engineering, where an open circuit or an unlocked door describes the state a device settles into when power is lost. In software the analogy is applied to any component that can be unreachable while the surrounding request continues: an authorization service, a content filter, a rate limiter, a feature flag provider, a logging pipeline, or a fraud check.

Failing closed is generally correct when the unavailable component enforces a boundary. If an authorization service cannot be reached, permitting the action is equivalent to removing access control while nobody is watching. Similar reasoning applies to payment verification, safety filters on agent output, spend limits, and anything whose failure mode is an unauthorized or irreversible action.

Failing open is generally correct when the component only observes or enriches. Telemetry, analytics, caching, recommendation ranking, and optional context retrieval should not take down the primary flow when they are down. A monitoring system that blocks requests when it cannot record them converts an observability incident into a customer facing outage.

The confusing part is that the words are used inconsistently. In security literature fail safe often means fail closed, while in availability literature fail safe can mean the system keeps serving. Because of this ambiguity, careful documentation states the concrete behavior, for example that requests are rejected with a specific error when the policy service times out, rather than relying on the label alone.

## Key points

- Fail open continues the operation; fail closed blocks it.
- Enforcement components should generally fail closed.
- Observation and enrichment components should generally fail open.
- The wrong default silently removes a control or causes an outage.
- Terminology is inconsistent, so state the concrete behavior.

## In practice

An agent platform runs two checks before a tool call. A spend guard verifies remaining budget, and a telemetry hook records the call. When the spend guard is unreachable the platform fails closed and refuses the call, since proceeding could run up unbounded usage nobody can see. When the telemetry hook is unreachable it fails open, logging a local warning and letting the call proceed, because losing a metric is far less costly than halting the work.

## Related terms

- [Graceful Degradation](/en/glossary/graceful-degradation)
- [Circuit Breaker](/en/glossary/circuit-breaker)
- [Monitoring](/en/glossary/monitoring)
- [Alerting](/en/glossary/alerting)
- [Observability](/en/glossary/observability)

[Back to the AI Glossary](/en/glossary)