WhiteFin is the execution layer of the AI agent stack — Layer 4 — the layer that governs the gap between what the agent decides and what the tool actually runs. It enforces at the kernel boundary using eBPF, intercepting every system call an agent makes before it completes, denies by default, and provides verifiable evidence that the action came from the agent's stated intent.

How is WhiteFin different from prompt filters or monitoring tools?

Prompt filters look at what the model says. Monitoring tools observe what already happened. WhiteFin enforces at the kernel boundary — it intercepts every system call before it completes and refuses the ones that are not explicitly approved. Observation is useful but it is not enforcement — agents do not slow down for monitors that cannot say no.

What is Causal Provenance?

Causal Provenance is the verifiable evidence that an agent's action actually came from its stated intent. Cryptographically signed. Tamper-evident. Independently verifiable without trusting WhiteFin.

How long does it take to get to enforcement?

Fifteen days. Two weeks of shadow-mode observation, one day of policy review, one day to flip the switch. You do not author the policies; we generate them from what your agents actually do, and you approve them before anything goes live.

Can WhiteFin be deployed air-gapped?

Yes. Two deployment options: self-hosted and air-gapped. The enforcement guarantees are identical across both; you pick by data-residency and audit posture.

PII Re-hydration: Zero-Trust Data Protection for LLMs

Gilad GabayJanuary 5, 20261 min read

How WhiteFin ensures your sensitive data never reaches external LLM providers. A technical look at tokenization and secure storage.

PII Re-hydration: Zero-Trust Data Protection for LLMs

One of the biggest concerns enterprises have with AI adoption is data privacy. WhiteFin PII Re-hydration system solves this with a zero-trust approach.

How It Works

1. PII Detection

We scan for 17 types of sensitive data:

Email addresses
Phone numbers
Credit card numbers
Social Security Numbers
Names and addresses
Medical IDs

2. Tokenization

Each detected PII entity is replaced with a secure token.

3. LLM Processing

The tokenized prompt is sent to the LLM - they never see real PII.

4. Re-hydration

When the response comes back, we replace tokens with original values.

Security Properties

Zero Trust: LLM providers never see real PII
Ephemeral Tokens: 5-minute TTL, then purged
Encrypted Storage: AES-GCM with per-client keys
Audit Trail: Every tokenization/rehydration is logged

#pii#privacy#zero-trust

Gilad Gabay

Co-Founder & Chief Architect

LinkedIn Email