Detects prompt injection by its effect on a sacrificial canary model, not just pattern matching: untrusted input hits a powerless model first, a behavioral check reads the residue, and it returns block, flag, or pass before your primary model acts. Inbound preflight sensor, not a guarantee.
Feature flags for Apache Airflow: ramp, A/B test, and kill-switch your DAGs' behavior and placement (pool/queue/executor) through OpenFeature. Works with flagd, LaunchDarkly, GrowthBook, Unleash, Statsig, or an in-house engine.
