Fail-Open

A dangerous failure mode where a system continues operating with weaker checks or broader permissions after a control breaks or becomes unavailable.

Fail-open describes the condition where a security control fails and the surrounding system responds by continuing operation in a more permissive mode. Instead of halting, downgrading safely, or escalating for review, the workflow keeps moving with fewer guarantees.

In traditional security, fail-open patterns appear when authentication outages still permit access, network controls silently allow traffic, or validation errors are ignored. In agentic systems, the pattern is especially dangerous because one fail-open decision can be replayed across multiple tools, memories, and downstream actions.

Fail-Open in Agentic Workflows

Common AI-specific fail-open patterns include:

missing provenance metadata treated as safe by default,
policy-engine timeouts converted into "best effort" execution,
failed simulation followed by direct transaction staging,
sandbox unavailability routed into unrestricted execution,
and incomplete tool descriptors accepted without risk classification.

These patterns are often introduced for reliability or user-experience reasons, but they expand blast radius dramatically. For high-impact actions, a secure design typically prefers fail-closed behavior or graceful degradation over silent continuation.