Verify, Inspect, Govern, Isolate, Prove, what each one does and where it lives in your existing perimeter. The full platform deploys behind a single routing rule and runs in your environment.
BladeRun verifies the agent, binds it to its mandate, enforces at the boundary, and proves what happened, then hands a verdict and signed evidence to your fraud, identity, and core systems. We don't replace any of them.
BladeRun is a verifier, not an issuer. We work with whatever identity your bank already uses, no parallel system to provision, no new directory to maintain. Every outbound agent call carries a signed credential that gets validated at the egress before anything else runs.
If a credential is missing or invalid, the call is rejected at the agent boundary before reaching your model providers. Your existing identity team owns the source of truth; we own the runtime check.
If your bank also exposes customer-facing AI surfaces (agent-accessible portals, public banking-agent flows), the same verification runs at the page layer through BladeRun.js.
The Gateway sits in the call path between your AI agents and any model provider, your existing one or the next one you add. It runs structural inspection on every outbound prompt and every returning response, including content fetched into the agent's context from external sources.
What gets inspected is calibrated to your traffic during a shadow-mode pilot before any enforcement is enabled. Nothing surprises your SOC.
Your governance team writes policy once. The Gateway enforces it on every outbound call: sensitive customer data is handled per your contract with each provider, and every agent operates within an explicit, signed authorization scope.
Scope is not only per-call. The Gateway enforces the size and shape of the mandate over time, holding an agent to aggregate spend, velocity, and cumulative caps rather than only checking whether a single call is in scope. A hundred $99 charges is not a $99 mandate. Generic per-call scope enforcement is what every vendor claims; holding an agent to the mandate in aggregate is the harder guarantee.
The same policy plane runs across all of your model providers, OpenAI, Anthropic, Bedrock, Azure, Vertex, or self-hosted. Adding a provider does not multiply your governance work.
Overseer runs alongside the Gateway, scoring every agent action against a behavioral baseline that matures during deployment. When something deviates, a single agent or a coordinated pattern across the fleet, the Kill Switch isolates the affected sessions automatically.
Containment is automatic and reversible; escalation to a human is mandatory by policy for defined action classes. Your SOC reviews the incident afterwards with the full forensic record. The detection plane is designed to be defensible under your existing model-risk-management process.
Time Machine writes the complete record of every agent session, every prompt, every action, every policy decision, to your S3 bucket, under your KMS keys, under your retention policy. Bank CISOs do not accept compliance evidence living in a model provider's infrastructure. Sovereignty is the point.
The record is signed and verifiable end-to-end. This is what you show examiners. This is what you replay after an incident. This is what you own regardless of which providers you use tomorrow.
Lives inside your agent. Captures every action and signs every session.
Explore Sentinel SDK →The egress reverse proxy. Verify, inspect, govern, redact, all at the boundary.
Explore Gateway →Per-agent and fleet-wide detection. Automatic isolation on anomaly.
Explore Overseer →Forensic record of every agent action in your storage under your keys.
Explore Time Machine →Point one workflow at the BladeRun Gateway. The first week runs in shadow mode. Your team sees the same capabilities running on real agent calls before any enforcement is enabled.
