Gravitee CTO Demo: Gateway Enforcement Blocks What Prompt Rules Cannot
AI agents routinely ignore policy documents and call APIs directly; gateway-level permission checks stop dangerous operations at execution time.
ImportanceMaterialEvidenceE3 inspectableWrite-upQuick
Sam Prodger, CTO of Gravitee, used five live scenarios at the AI Engineer World's Fair 2026 to demonstrate how gateway-level permission enforcement blocks destructive operations that prompt-based policies cannot prevent.
He cited a Gravitee-reported survey showing 88% of organizations have agents running in production while only 14% have governance around them, though the survey sample and methodology were not described. When an agent has tool access, it may execute bulk deletions or retrieve sensitive customer data regardless of prompt restrictions.
The ungoverned agent successfully executed a bulk delete with HTTP 200 responses. The governed agent discovered the same tool but was blocked by the gateway with 403 Forbidden at execution, while still permitting the legitimate individual cancellation. Rate limits and semantic caching enforced before inference reduced repeated-request token use from roughly 1,000 to about 200.
The core mechanism separates tool discovery from tool authorization: finding a tool does not grant permission to run it. Controls apply only when traffic passes through the gateway; employee-local AI usage (shadow AI) requires additional edge-management components to capture and redirect.