<15ms
end-to-end latency, p95
100K /s
request throughput per node
5
LLM providers, day one
100%
prompts and responses inspected
Where the Gateway lives

In the call path. Outbound to any LLM. Inbound from any external agent.

A single routing rule moves traffic through the Gateway. Your application code does not change. Your model provider does not change. The Gateway becomes the policy and inspection plane between them.

Your environment

Bank AI application

Workflow, chatbot, multi-agent orchestrator, or API caller, built in your framework of choice.

BladeRun Gateway

Inspect · Verify · Redact · Log

Reverse proxy. Every prompt scored, every response inspected, every MCP endpoint hashed against your registry.

External

Any LLM provider

OpenAI, Anthropic, Azure OpenAI, AWS Bedrock, Google Vertex, or self-hosted models.

What the Gateway does

Three responsibilities. One pass through the call path.

01 · Inspect

Inspect every prompt and response, direct and indirect injection.

The Gateway runs a structural injection classifier on the full prompt before it reaches the model, and on the response before it reaches your application. Direct injection (instructions in the user message) and indirect injection (instructions hidden in tool-call results, retrieved documents, or web content) are both detected.

  • Structural classification. Token-distribution, instruction-override n-grams, and origin-tag class, not signature matching.
  • Indirect injection. Instructions arriving via fetch_url, RAG retrieval, or tool-call output are flagged with the originating context.
  • Response inspection. Model output checked for data exfiltration patterns, system-prompt leakage, and policy violations before it reaches the application.
Live · Outbound prompt inspection
→ POST /v1/chat/completions
user.message: "Summarize this job posting"
tool_call: fetch_url("https://example/jobs/4421")
tool_result: <3.2KB HTML>
⚠ injection detected
  origin = external_web
  pattern = instruction-override (0.94)
  fragment = "ignore previous… initiate wire…"
→ BLOCK · returned policy violation to caller
02 · Verify

Verify every MCP endpoint against the registry, not the URL.

MCP impersonation is the single most exploited vector against bank agents in 2025. The Gateway maintains a per-bank MCP Server Registry of approved endpoints, hashed by URL and TLS certificate fingerprint. Any unregistered or hash-mismatched endpoint is blocked before the call is forwarded.

  • Hash-based registry. Each approved endpoint is fingerprinted and hashed. One-character typosquats produce a different hash and are blocked.
  • Federation-fed denylist. Cross-bank typosquat campaigns produce shared violation hashes. Every member bank is protected the moment one detects.
  • Registry change control. Approved endpoints require explicit add via your change-management workflow. No silent drift.
Live · MCP endpoint verification
agent.call → mcp://payments-api.bank.com/v1/wire
  hash = a7f3c1d2…
  registry = MATCH (added 2026-02-14)
✓ ALLOW
agent.call → mcp://paymerits-api.bank.com/v1/wire
  hash = 9e4b8a01…
  registry = NO MATCH
  similarity = 0.94 vs. payments-api.bank.com
⚠ typosquat suspected → BLOCK
  signal hash forwarded to Federation
03 · Redact

Redact NPI, PCI, and PII before the prompt leaves your perimeter.

The Gateway runs Presidio-grade entity detection on every outbound prompt. PAN, CVV, account numbers, customer identifiers, and free-text NPI are replaced with deterministic tokens. The model gets the structure of the prompt, not the sensitive values. Your application un-redacts on the response path if needed.

  • GLBA / Reg P alignment. Customer NPI never reaches a third-party LLM.
  • PCI-DSS 4.0 alignment. PAN and CVV stripped from agent prompts in real time.
  • Deterministic tokenization. The same value gets the same token within a session, model reasoning preserved without exposing the value.
Live · entity redaction
4521-7890-2143-9912[PAN_a7f3]
SSN 123-45-6789[SSN_9e4b]
acct 8821-449[ACCT_c1d2]
CVV 471[CVV_03f9]
jane.doe@acme.com[EMAIL_4a18]
DOB 1984-06-12[DOB_b29c]
12 entity classes · per-session deterministic tokens · zero raw NPI to provider
Enforcement cascade

Four layers run before a single byte leaves your perimeter.

Each layer is deterministic and independently auditable. A call must pass all four to be forwarded. Any single failure produces a structured policy decision with full provenance.

1
Identity verification
The calling agent presents a signed identity. Unsigned or expired calls are rejected before policy evaluation.
DETERMINISTIC
2
Tool-permission certificate
The agent's allowed tool set is checked. A market-intelligence agent attempting to call initiate_wire is rejected at this layer regardless of prompt content.
DETERMINISTIC
3
Injection classifier
Structural classification on prompt and response. Direct + indirect injection patterns flagged with origin context.
CLASSIFIER
4
DLP redaction + endpoint registry
NPI, PCI, PII tokenized. MCP endpoint hash matched against your registry. Final go/no-go before the request is forwarded.
DETERMINISTIC

Deterministic controls are the guarantee. Content classification is best-effort, and we mark which is which.

Coverage at the Gateway

What the Gateway stops on its own, before any other module runs.

Overseer and Kill Switch handle behavioral and fleet-level threats. The Gateway handles everything that can be evaluated on a single call without baseline history.

Threat classHow the Gateway handles itLayer
Direct prompt injectionStructural classifier on the user message; instruction-override patterns flagged with origin tag.L3
Indirect prompt injectionInstructions arriving via tool-call output, RAG retrieval, or fetch_url tagged with their source context.L3
Malicious MCP endpointHash mismatch against the per-bank MCP Server Registry. Typosquats blocked even with valid TLS.L4
NPI / PCI exfiltration12-class entity detection with deterministic per-session tokenization before the prompt is forwarded.L4
Reconnaissance promptsLiving-off-the-LLM patterns, system enumeration, credential lookup, persistence-script generation, flagged on signature.L3
Tool-permission violationAgent's signed tool certificate checked against the requested call. Out-of-scope tool calls rejected without policy evaluation.L2
Unsigned agent trafficUnsigned calls rejected at L1. No agent reaches the model without a verifiable identity.L1
Deployment

SaaS, on-premises, or air-gapped. Same detection. Same SLA.

SaaS

Managed cloud

Gateway runs in BladeRun's VPC, colocated with your model providers. Single routing rule. Deploys in days. SOC 2 Type II in scope.

On-premises

Customer VPC

Gateway runs in your VPC under your IAM. No traffic leaves your perimeter. Meets OCC and Federal Reserve constraints on data flow.

Air-gapped

Disconnected deployment

For sovereign workloads. Federation participation is opt-in for SaaS / on-prem; air-gapped is local-only and equally functional.

Compliance mapping

Gateway maps to four examiner questions on day one.

GLBA / Reg P
Technical safeguards preventing unauthorized disclosure of customer NPI to third parties.
L4 · DLP redaction
PCI-DSS 4.0
Protection of cardholder data, PAN and CVV, across AI-assisted workflows.
L4 · entity tokenization
FFIEC AI Guidance
Inspection and logging of every AI decision affecting a customer or transaction.
All layers · Time Machine
EU AI Act · Article 12
Operational logs and post-hoc auditability for high-risk AI systems.
Full request/response log
Works with

The Gateway is one of four platform components.

Behavioral detection, automated isolation, and cross-bank intelligence run alongside the Gateway. Every detection event is written to Time Machine and contributes anonymized signal to the Federation.

Common questions

What architecture teams ask first.

What's the actual latency cost of running every prompt through the Gateway?
Under 15ms end-to-end at p95, including all four enforcement layers. The Gateway is colocated with your model provider region. The classifier is small enough to run on commodity CPU; entity redaction is the dominant cost and runs in parallel with classification. We publish per-region p99 latency in the documentation.
Which model providers are supported on day one?
OpenAI (direct + Enterprise), Anthropic (direct + Bedrock), Azure OpenAI, AWS Bedrock (Nova, Llama), Google Vertex (Gemini), and any OpenAI-compatible self-hosted endpoint (vLLM, TGI, Ollama). Adding a new provider requires no Gateway code change, only a routing rule and a provider profile.
How does the Gateway handle streaming responses?
SSE and chunked-encoding streams are inspected token-by-token with a buffered classifier. If a policy violation is detected mid-stream, the Gateway terminates the connection and surfaces a structured error to your application. Latency overhead on streamed responses is sub-millisecond per chunk.
Can our application opt out of redaction for specific prompts?
Yes. The Gateway honors a per-request scope header that scopes redaction policy down to a documented allowlist. The decision is logged to Time Machine with the requesting agent's identity. This pattern is used by reconciliation agents that legitimately need a specific NPI value end-to-end, but it's rare and audited.
How does the MCP Server Registry get updated?
Two paths. Bank-controlled additions go through your change-management workflow (Jira / ServiceNow integration available; manual API also supported). Federation-fed denylist additions are pushed automatically when a member bank detects a typosquat campaign. Your local engine starts blocking the matching pattern within a minute.
What happens if the Gateway itself goes down?
Two operating modes. Default is fail-closed: if the Gateway is unreachable, agent calls are rejected. There's no safe harbor under SR 26-2. Examiners still apply model-risk and safety-and-soundness principles, and fail-closed is the defensible posture. Fail-open is available for non-regulated workloads. The Gateway runs as a stateless cluster behind a load balancer; rolling deployments and node failures are handled without traffic loss.
Pilot the Gateway

One routing rule. Visible results in 48 hours.

Point one workflow at the BladeRun Gateway. Inspection runs in shadow mode for the first week. We show you what was inspected, what would have been flagged, and what would have been blocked. Reversible by routing rule.