12

entity classes detected on day one

<5ms

redaction latency, p95

0

raw NPI ever forwarded to a model provider

GLBA·PCI

aligned by design

How it works

Detect, tokenize, un-tokenize on the response.

01 / Detect

Presidio-grade entity detection on every outbound prompt.

The DLP Engine scans the full prompt, including tool-call parameters, RAG retrieval results, and document attachments, for the 12 entity classes documented below. Detection runs on commodity CPU and parallelizes with the injection classifier; under 5ms at p95.

  • Pattern + ML hybrid. Regex for high-confidence formats (PAN, SSN); fine-tuned NER for free-text NPI.
  • Document attachments. PDFs and Office docs uploaded into agent context are redacted at attachment time, not at inference time.
  • RAG retrieval. Documents pulled from a vector store are redacted as they enter the context window.
Live · prompt entering DLP
Summarize the wire transfer from
account 8821-449 to recipient John Doe
at 555 Main St, SSN 123-45-6789, on
card ending 4521-7890-2143-9912.
DETECT
5 entities flagged
  acct, name, address, ssn, pan
Tokenized prompt, what reaches the model
Summarize the wire transfer from
account [ACCT_a7f3] to recipient [NAME_9e4b]
at [ADDR_c1d2], SSN [SSN_03f9], on
card ending [PAN_4a18].
FORWARD to model provider
model never sees the originals
02 / Tokenize

Deterministic per-session tokens. Model reasoning preserved.

Every detected entity is replaced with a token shaped like [ACCT_a7f3]. The same value gets the same token within the session, so the model can reason about "this account" and "this recipient" without ever seeing the underlying value. Tokens are scoped to the session and never reused across sessions.

  • Deterministic within session. The model sees consistent references; reasoning is preserved.
  • Salted across sessions. A token from yesterday's session is meaningless to today's. No cross-session correlation possible.
  • Format-preserving where useful. PANs are replaced with valid-looking but tokenized PANs to maintain prompt structure for tools that validate format.
03 / Un-tokenize

On the response path, tokens are restored, selectively.

If the model returns a response that references [ACCT_a7f3], the DLP Engine restores the original value before the response reaches your application. If your application does not need the un-tokenized value, you can opt into receiving the tokens directly. Per-application policy.

  • Per-application scope. Some downstream applications need the original value; some explicitly should not. Both patterns supported.
  • Audit trail. Every un-tokenization event is logged to Time Machine with the requesting application's identity.
  • Token leak detection. If a token appears anywhere it shouldn't, the DLP Engine flags the leak.
Response path, selective restoration
Model response:
"Wire from [ACCT_a7f3] to [NAME_9e4b]
 for $50,000 was processed at 14:22 UTC."
DLP restores tokens for caller
"Wire from acct 8821-449 to John Doe
 for $50,000 was processed at 14:22 UTC."
restoration logged · scope: payment-orch
Entity classes

12 entity classes shipping today. Bank-specific extensions on request.

The 12 classes below cover the regulated data categories that trigger compliance findings. Bank-specific extensions, internal account number formats, custom customer ID schemes, branch codes, are added through configuration and shipped with your deployment.

PCI
PAN, primary account number
CARD_PAN · format-preserving
PCI
CVV / CVC
CARD_CVV · removed entirely
GLBA
Bank account number
ACCT · per-bank format support
GLBA
SSN / Tax ID
SSN_TIN · pattern + context
GLBA
Customer name
NAME · NER + roster lookup
GLBA
Date of birth
DOB · context-aware
GDPR
Email address
EMAIL · domain whitelisting
GDPR
Phone number
PHONE · regional formats
GDPR
Postal address
ADDR · multi-line aware
GDPR
IP address
IP · v4 / v6
HIPAA
Medical record number
MRN · for healthcare workloads
GLBA
Free-text NPI
FREE_NPI · ML-based

Each entity class has documented detection precision and recall measured against a labeled dataset of bank prompts. The metrics ship with the deployment and update with each detector revision. Pattern-based classes are the guarantee; free-text NPI is best-effort content inspection, and we mark which is which.

Scope override

When you legitimately need a value end-to-end.

Reconciliation agents, fraud-investigation agents, and certain customer-service workflows legitimately need a specific NPI value to reach the model. The DLP Engine honors a per-request scope header that scopes redaction down to a documented allowlist. The decision is logged to Time Machine with the requesting agent's identity, the scope, and the human principal.

Default

Full redaction · 12 classes

The default for every agent. NPI never reaches the model. Most banks operate the entire fleet under this default and only deviate for specifically scoped agents.

Scoped

Per-agent allowlist

Specific entity classes can be allowed for specific agent types under documented scope. Scope changes go through your change-management workflow. Every scoped agent action is logged with its scope and human principal.

Compliance mapping

DLP answers the data-protection examiner question.

FrameworkWhat it expectsBladeRun control
GLBA / Reg PTechnical safeguards preventing unauthorized disclosure of customer NPI to third parties.12-class detection · tokenization
PCI-DSS 4.0Protection of cardholder data, PAN, CVV, across all processing environments.PAN format-preserving · CVV removed
GDPR · Article 32Pseudonymization and encryption of personal data.Per-session deterministic tokens
HIPAA · 164.312Technical safeguards preventing unauthorized disclosure of protected health information.MRN class · scoped extensions
FFIEC AI GuidanceDocumented controls for data flowing into and out of AI systems.Time Machine · all redactions logged
Works with

DLP runs as Layer 4 of the Gateway enforcement cascade.

L4 · Gateway

Gateway

DLP is the final layer of the Gateway enforcement cascade, runs on every outbound prompt before forwarding.

Explore Gateway →

Prove

Time Machine

Every redaction is logged with the original entity class, the token, and the agent's identity. Restoration events are logged separately.

Explore Time Machine →

Federate

Federation Network

Detector improvements ship through Federation rule updates, when one bank improves a detector, every member benefits.

Explore Federation →

Common questions

What privacy and compliance teams ask first.

What's the false-positive rate on free-text NPI detection?

Pattern-based detectors (PAN, SSN, account number) operate at well above 99% precision because the underlying formats are structured. Free-text NPI is the harder class; we publish per-deployment precision and recall against your bank's actual prompt corpus during the pilot. Tuning happens during the maturation phase before enforcement is enabled.

Does redaction degrade model output quality?

Less than most teams expect. Deterministic per-session tokens preserve the model's ability to reason about entities ("this account," "this recipient") and produce coherent responses. We see no measurable degradation on benchmarks where the model is asked to reason about structure, summarize, classify, or extract. For workflows where the model needs the underlying value, the per-agent scope override pattern applies.

How do we add bank-specific entity classes?

Configuration. Internal account number formats, custom customer ID schemes, branch codes, and similar formats are added by submitting a regex or sample dataset. The detector ships with your deployment. Bank-specific detectors do not contribute to the Federation Network. They remain local.

What happens to attachments in agent context, PDFs, spreadsheets?

Attachments are redacted at the point they enter the agent's context, not at inference time. PDFs are extracted to text, redacted, and re-attached as redacted text or as a sanitized PDF depending on the workflow. Office docs are handled the same way. The original attachment never reaches the model.

How do we handle RAG retrieval results that contain NPI?

RAG retrieval results pass through DLP as they enter the context window. The model sees redacted retrieval. Two configurations are supported: redact-then-rerank (the model never sees raw NPI) and rerank-then-redact (rerank quality preserved on the original; redaction applied just before forwarding). The first is the default for regulated workloads.

Where are the deterministic tokens stored?

In your environment. The token-to-value mapping lives in a per-session encrypted store inside the Gateway, scoped to the session lifetime. Long-lived sessions can configure a longer scope window. The mapping is never persisted outside your environment, never transmitted, never accessible to BladeRun staff.

Pilot the DLP Engine

Run it in shadow mode. See the leakage before you decide.

The first week of every Gateway pilot includes DLP in shadow mode. We report what would have been redacted without modifying the prompts. Most teams are surprised by what shows up.