Data privacy and protection with verified outputs for AI agents
Control what your AI agents can see, hold, spend and claim, without rebuilding your agent stack.
Built on The Vault, NoPII and Triplets.
- PCI DSS Level 1
- SOC 2 Type II
- AWS Foundational Technical Review
- Sub-100ms vault operations
- 99.99% uptime SLA
AI agents create new paths for sensitive information.
An agent may retrieve a customer record, assemble context from several systems, invoke a payment service and update a case, all within one task. Sensitive values can travel far beyond the database in which they began.
For security and technology leaders, the question is no longer only who can access a system. It's what information the agent can retrieve, where that information travels and what the agent is allowed to do with it.
Access can exceed the task
Broad connector credentials may expose records or files unrelated to the request.
Information spreads through the stack
Prompts, tool results, traces, memory, caches and vector stores can retain copies of sensitive information.
Outputs become business decisions
An answer may look convincing despite missing evidence, outdated policies or conflicting source documents.
Actual exposure depends on framework configuration, connectors, logging and retention settings.
Protect the data before the agent uses it. Verify what it produces.
AI Vault places Enigma's existing data protection and verification technologies into the paths an agent uses to retrieve information, call models, act through tools and respond. The agent works with permitted data and references; access to protected values and release to business systems remain subject to controls outside the model.
Enterprise information
Customer records
Cards & payments
Documents & files
Protect and tokenize
Store sensitive values behind a defined boundary. Pass tokens or references into supported model paths.
AI agent / LLM
Retrieves, reasons and acts using task-relevant tokens and references.
Scope access
Determine which records, files, tools and actions an agent workflow is authorized to use.
Minimize exposure
Tokenize sensitive fields before supported model calls and limit unnecessary copies downstream.
Govern release
Keep original values inside the protected boundary until an approved destination requires them.
Check the output
Evaluate eligible response claims against approved evidence and apply the workflow's handling policy.
Three established technologies, extended for AI agents.
AI Vault is an integrated application of The Vault, NoPII and Triplets, not a replacement for them. Each technology addresses a different point of exposure or uncertainty in an agent-driven workflow.
The Vault
Protect information at the source. Release only what the workflow permits.
Customer records, payment credentials and confidential files do not need to become part of an agent's general working context. The Vault supplies protected storage, tokenization and controlled data handling that can be connected to task-specific access policies.
For payment operations, an agent can initiate a workflow using a card token while a supported vault proxy handles the original value at the gateway boundary. For records and documents, access can be scoped to the permitted information rather than an entire repository.
- Protected customer data, cards and files
- Task-specific access and controlled release
- Vault access records and payment proxy paths
NoPII
Keep raw personal information out of the model path.
Data privacy for AI must cover more than the initial prompt. Agents repeatedly retrieve information, compose context and exchange results with models and tools. NoPII identifies and tokenizes supported sensitive values before they are sent to the model, preserving useful references without unnecessarily exposing the originals.
Within the enforced data path, logs, traces and other downstream systems can observe tokens rather than raw values. The original information is restored only through an authorized return or release path.
- PII detection and tokenization before model requests
- Deterministic references through supported interactions
- Controlled restoration when an authorized application needs the value
Member: 4471-2290
Email: maria@example.com
Member: [ID_1]
Email: [EMAIL_1]
Triplets
Verify what the evidence supports, not just what the model says.
A coherent answer is not necessarily a defensible one. Triplets evaluates eligible claims in an agent's response against an approved, versioned evidence corpus, including policies, contracts, documentation and internal records.
It identifies claims that are supported, refused because evidence is insufficient, or conflicted because sources disagree. Verdicts retain source and corpus-version references so reviewers can understand what the answer relied on and reassess it when the evidence changes.
- Claim-level evidence checks, beyond attached citations
- Supported, refused and conflicted verdicts
- Traceability to source material and its version
A compromised agent should not become a data breach.
Prompt injection or a manipulated tool response can cause an agent to attempt an unauthorized action. The impact depends on whether the model has direct access to the sensitive values behind its tools.
AI Vault is designed to separate the agent's working context from protected values. A token or reference does not automatically grant the ability to retrieve a card number, a customer record or a confidential file.
The protection applies to integrated, enforced data paths. It reduces the consequences of a compromised agent; it does not remove all prompt-injection or tool-security risks.
Make sensitive agent workflows safer to operate.
AI Vault is designed around the systems, data and decisions involved in each use case. These examples illustrate how The Vault, NoPII and Triplets can work together.
Insurance claims and coverage
An agent retrieves a policy, claim documents and relevant exclusions to explain coverage to a customer.
Built around your existing AI environment.
AI Vault is designed to work with your agent framework, model providers, enterprise APIs, document stores and business systems. It introduces protection and evidence checks at the points where information is retrieved, processed, released and returned.
We start by mapping one workflow: its identities and permissions, the data it touches, the tools it can invoke, the destinations that need original values and the evidence that should govern its answers.
- Align access decisions with existing identity and application policy
- Define protected data paths, release rules and audit requirements
- Evaluate tokenization coverage, evidence handling and operational fit
The Vault
NoPII
Triplets
Questions enterprise teams ask.
Is AI Vault another AI agent platform?
No. AI Vault is a data protection and output verification layer for the agents you already run. It brings The Vault, NoPII and Triplets together around your existing models, frameworks and business systems. It does not build, host or orchestrate agents.
How is AI Vault different from a data privacy vault?
A data privacy vault stores sensitive values and returns tokens. AI Vault includes that vault, through The Vault for records, cards and files, and extends it into the agent's working path. NoPII tokenizes personal information before supported model calls, release decisions are evaluated outside the model, and Triplets checks the agent's answers against approved evidence.
How is NoPII different from PII redaction or masking for LLMs?
Redaction removes values from a prompt. NoPII replaces them with deterministic tokens, so the model can still reason about the same person or record across an interaction, and an authorized application can restore the original value. AI Vault also covers what redaction does not: the cards and files agents act on, and whether the answer is supported by your evidence.
Does the agent ever need to see the original sensitive value?
In many workflows it can operate on tokens and references. If an approved application or gateway needs an original value, the workflow routes that value through a controlled release path. The exact boundary depends on the integration and the authorized destination.
How do agents use payment cards without holding card numbers?
The agent holds a token. When a payment is approved, the vault releases the card to the processor through an HMAC-signed proxy on supported gateways. CVV values are purged within 30 minutes.
Which model providers and data types does NoPII cover?
NoPII works with nine model providers and detects personal information along with 14 families of secret patterns, such as keys and credentials. If a protected request cannot be tokenized, it fails closed rather than sending raw values to the model. See the NoPII documentation for the current provider list.
Can you govern access to individual files, records or payment operations?
AI Vault can be designed to enforce task-specific access and release policies through The Vault and the connected systems. The required identity, authorization and approval mechanisms are defined for your environment rather than assumed to be the same across every agent framework.
What happens if an agent is prompt-injected or manipulated?
On protected paths, the agent holds a reference without automatically gaining access to the underlying value. Requests to use protected information are checked outside the model and logged. This limits exposure but does not replace other defenses for agents, tools and enterprise infrastructure.
What does a verified output mean?
Triplets checks eligible claims in an agent's response against an approved, versioned evidence corpus and labels each one supported, refused or conflicted, with links to the specific sources and versions. A verified output is one whose claims are supported by the evidence on file. It is not a guarantee of truth beyond that evidence. Your organization decides whether refused or conflicted content is held, revised or reviewed before release.
Can AI Vault run in our own environment, and how do we buy it?
NoPII and Data Vault are available as self-hosted containers, and other deployment options are scoped per integration. Enigma Vault is available on AWS Marketplace and through direct enterprise agreements.
How do we evaluate an AI Vault integration?
Start with one agent workflow. We review the data and tool paths, agree on protection and evidence requirements, integrate the controls into your existing stack, and evaluate coverage, authorization, source traceability, latency and operational fit.
Protect the data your agents use. Verify the answers they give.
Tell us what your agent needs to access, which systems it works with and where verification matters. We'll help define the right protection and evidence controls for your workflow.
Discuss your agent workflow Request a no-cost POC