Behavioral Canarying for Prompt Injection: Powerless Model Probes with Explicit Coverage Semantics

Behavioral canarying is a pre-execution prompt-injection sensing architecture: untrusted input is deliberately exposed to a powerless language-model probe, and its response is examined for evidence of behavioral compromise before an authority-bearing agent acts. Little Canary treats that response as non-authoritative evidence. Its central implementation contract keeps routing disposition separate from inspection coverage, distinguishing an input allowed after completed inspection from one allowed because inspection was unavailable or degraded under a fail-open policy. This technical note describes the architecture, its evidence boundary, and its relation to classifier-, token-, and containment-based approaches. It does not claim universal detection, formal security, aggregate accuracy for the current release, or invention of the general sacrificial-canary concept.