Skip to main content

Data Protection for AI Agents

Data protection for AI agents, on every prompt and tool call

Redact PII, PHI, financial data and secrets from agent and LLM traffic before an agent or model sees them, with rules that follow each person, agent and team.

Data protection for AI agents is the set of controls that keep sensitive data out of what agents and models see and send: redaction of PII, PHI and secrets in prompts and tool responses, decided on every call for the person and agent involved.

Agents are useful because they reach real data: customer records, tickets, code and patient files, and that reach is also the exposure. Every prompt can carry customer details to a model provider, and every tool response can hand an agent more than its task needs. SecureAuth Agent Authority treats AI access control and data security as one decision, so the rule that lets an agent call a tool also decides what data comes back.

  • Customer details removed from prompts before they leave your company
  • PII and PHI redacted from tool responses before the agent reads them
  • Every call recorded, with redactions counted and the data never stored

The problem

Sensitive data leaks through AI agents in both directions

Agentic AI data protection has to cover more than one exit. Data leaves in prompts, arrives in tool responses and hides in code and documents.

  • Prompts carry customer data out

    People paste customer names, phone numbers and account details into prompts every day, and the prompt goes to a model provider outside your company.

  • Tools send back more than the agent asked for

    A lookup returns the whole record: dates of birth, health details, salaries. Once it is in the agent's context, it can end up in a summary, a ticket or the next prompt.

  • Code reads expose secrets

    A coding agent that reads a repository also reads the cloud keys and tokens someone committed to it, and passes them to the model behind it.

  • Tool output carries hidden instructions

    A document, ticket or web page can hide instructions aimed at the agent, such as asking it to ignore its rules or send data somewhere else.

  • One global filter fits few teams

    A single DLP rule set either blocks the people who need the data or lets everyone else see it. Support, clinical staff and contractors need different answers.

  • No record of what was removed

    When an auditor asks what an agent saw, logs that show only an API key can't say which agent, which person or which rule was involved.

DLP for AI agents covers prompts out and responses in

Agent traffic runs two ways, and each has its own checkpoint. Both apply the same identities and the same policy.

Prompts to models

LLM Authority finds customer names, phone numbers and account details in a prompt and removes them before the prompt leaves, so ChatGPT, Claude or Gemini receives the same request with placeholders. PII redaction for LLMs gives you two options: strip the details and let the work go on, or stop a prompt that shouldn't leave at all.

Responses from tools

MCP Authority filters what MCP servers send back before it reaches the agent. Apps often return more than the agent asked for; the response filter removes what this person and this agent should not see.

Policy example

A PHI redaction policy for one support team

A redaction policy picks the responses it applies to the same way a rule picks calls: by server, tool, people, groups and agent, with an exception written as a condition. The example below follows one policy from start to finish.

  1. Pick the responsesResponses from the get_patient and search_visits tools on the patient-records MCP server, for anyone in the Support group, through any agent.
  2. Write the exceptionUnless the person is in clinical-staff, who need the full record. The condition is validated and tested before it goes live.
  3. Filter structured fieldsPattern filters replace dates of birth, Medicare Beneficiary Identifiers and Social Security numbers with labels.
  4. Filter free textAn AI judge removes diagnoses, medications and test results from notes, where no pattern can find them. If the judge fails, the response is blocked.
  5. Record the outcomeThe call is allowed with redaction and written to Agent Trail with a count of what was removed.

The support agent still gets what the job needs, the patient and the next visit, so the work goes on. The policy works the same way outside health data: swap the server, the group and the filters and the same shape redacts card numbers for a contractor team or salary data for analysts.

Redaction presets for PII, PHI, financial data and secrets

Start from presets instead of writing patterns by hand. They work in structured fields and free text alike, and go beyond regex patterns where the data has no fixed format.

  • Pattern filters catch exact formats such as card numbers, routing numbers and dates of birth
  • An AI judge follows a plain-language instruction for what a pattern can't describe
  • You choose what happens if the judge fails, such as blocking the response
A response filter with its preset library open in the Agent Authority console: patterns grouped into personal data, financial, credentials and secrets, and network and infrastructure, such as email addresses, US Social Security numbers, card numbers, API keys, AWS access key IDs, JWTs, phone numbers and UK national insurance numbers, each shown with the pattern it matches.
Data Protection · Filter presets, with sample data

Personal data (PII)

  • Email addresses
  • Phone numbers
  • US SSNs
  • UK NI numbers

Personally identifiable information, in fields and in free text.

Health data (PHI)

  • Dates of birth
  • Medicare Beneficiary Identifiers
  • Diagnoses and medications

Protected health information, including what a pattern can't describe.

Financial data

  • Card numbers
  • Routing numbers
  • Salary and compensation

Payment and pay data, removed before an agent or model sees it.

Credentials and secrets

  • API keys
  • AWS access key IDs
  • GitHub and GitLab tokens
  • JWTs

Keys and tokens stopped from leaving in prompts or tool calls.

Network and infrastructure

  • Network addresses
  • Infrastructure details

Details of your environment kept out of model and agent context.

Data controls follow each person, agent and team

Data controls are applied per identity, so each group can have its own rules. The same server can return different data to different people, and to the same person through different agents.

Rules use the groups and attributes already in your identity provider, such as Okta or Microsoft Entra ID. Finance and support can have stricter rules than engineering, and a rule can name a specific agent, such as Claude Code, for one group of analysts.

Because each agent has its own identity, tied to the person it acts for, the decision knows who is asking and why. For how those identities are issued and authorized, see AI agent identity and authorization.

Redact is one of five decisions on every call

Redaction is not a separate product bolted on after access. Each call gets one decision, and redacting is one of the five it can reach.

  1. AllowWithin policy, the call proceeds as is
  2. RedactThe call proceeds with PII, PHI and secrets removed
  3. ThrottleSlowed to the agent's or person's limits
  4. Ask a personHeld for approval with CIBA, on the person's own device
  5. DenyBlocked, and recorded with the rule that blocked it

In practice, a blocked call stops the agent, while a redacted call lets it finish the task without the data it shouldn't have. When the data is too sensitive to send even with redaction, the same rule can deny the call or hold it for a person's approval instead.

Strip credentials from code and neutralize prompt injection

Two response filters that protect the agent as much as the data: one removes secrets before the agent reads them, the other removes instructions an attacker hid for it.

Credential stripping

On your code-hosting MCP servers, file and code-search responses pass through patterns that replace AWS access keys, GitHub tokens and GitLab tokens with redaction labels before a coding agent can see them. More on AI coding agent security.

A response filter that strips cloud credentials from code reads, open in the Agent Authority console: on two code-hosting MCP servers, responses from file and code search tools pass through patterns that replace AWS access keys, GitHub tokens and GitLab tokens with redaction labels before a coding agent can see them.
Data Protection · Credential filter, with sample data

Prompt injection defense

Instructions hidden in documents, tools or web content are caught before the agent acts on them. An AI judge redacts text aimed at an AI assistant, such as requests to ignore previous instructions or reveal the system prompt, and blocks the response if the judge fails.

A response filter that neutralizes prompt injection in tool output, open in the Agent Authority console: for any agent, user and MCP server, an AI judge is instructed to redact instructions aimed at an AI assistant that are hidden in tool output, such as requests to ignore previous instructions or reveal the system prompt, and to block the response if the judge fails.
Data Protection · AI judge, with sample data

Keep the DLP and AI guardrails you already run

You don't have to replace the data protection tools your security team has tuned. Agent Authority puts them on the agent's call path, in the same decision as your policy.

  • Third-party data redaction

    Works with the data redaction and DLP services you already run, alongside the built-in filters.

  • Third-party guardrails

    Works with third-party AI guardrails, so their checks sit in the same decision as your policy.

  • Checks on your own model

    Content checks can run on a model endpoint you already approve, so prompts are inspected where you choose.

Every redaction recorded in Agent Trail, without the data itself

An audit trail that copies the sensitive data it removed is a new place to leak it. Agent Trail records that a redaction happened and leaves the redacted text out.

Every call is written to Agent Trail in order: the agent, the person it acted for, the MCP server or model, the decision, the rule that made it and any redactions. Redactions are recorded as counts only, so the matched text is never stored.

  • Search by agent, person or tool to see what ran and what was removed
  • See how allowed, denied and redacted decisions trend across agents, tools and teams
  • Stream decisions to the SIEM, SOAR and XDR tools your security team already runs

AI data governance

AI data governance starts with what agents can see

AI data governance is the practice of deciding which data AI models and agents may use, enforcing it on every use and recording it. Data governance for AI has to run at agent speed.

Regulations such as HIPAA for health data, GDPR for personal data in Europe and PCI DSS for card data are why many teams can't roll out agents until they can answer three questions: what data can each agent reach, what leaves in a prompt and who saw what. To answer them, the policy has to be enforced on each call.

Agent Authority gives your governance program the controls to point to: personal, health and payment data redacted per person and agent, sensitive calls held for approval and a record of every decision. It does not certify compliance on its own; your compliance team decides how these controls fit your obligations. To find and own the agents themselves, see agentic AI governance.

Evaluate

What to look for in an AI data security platform

Most AI data security tools apply one filter to all traffic. Compare that with data controls decided per identity, on every call.

Agent Authority compared with a single global DLP filter for AI agent traffic
Redaction rules differ by person, agent and teamA global filter applies one rule set to everyone, whoever is asking and why.Agent AuthorityYesGlobal DLP filterNo
Sensitive data found in free text as well as fixed formatsAn AI judge handles what a pattern can't describe, alongside pattern filters.Agent AuthorityYesGlobal DLP filterRegex patterns
Covers prompts to models and responses from toolsBoth directions of agent traffic, through the LLM and MCP gateways.Agent AuthorityYesGlobal DLP filterPartial
Credentials stripped from code readsBefore a coding agent, or the model behind it, can see them.Agent AuthorityYesGlobal DLP filterNo
Prompt injection in tool output neutralizedAgent AuthorityYesGlobal DLP filterNo
Redact is one of five decisions on the same callAllow, Redact, Throttle, Ask a person or Deny, decided together with access.Agent AuthorityYesGlobal DLP filterPass or block
Works with the DLP and guardrails you already runAgent AuthorityYesGlobal DLP filterPartial
Each redaction recorded with the agent, person and ruleRecorded as a count, so the audit trail never stores the matched text.Agent AuthorityYesGlobal DLP filterPartial

Data protection for AI agents FAQ

What is data protection for AI agents?

Data protection for AI agents is the set of controls that keep sensitive data out of what AI agents and models see and send. It redacts PII, PHI, financial data and secrets from prompts before they leave and from tool responses before an agent reads them, strips credentials from code, neutralizes prompt injection and records every call. In SecureAuth Agent Authority these controls are applied per person, agent and team, as part of the same decision that authorizes the call.

What is AI data governance, and how is it different from data governance?

Data governance sets the rules for who may use which data and how it is classified, stored and retained. AI data governance applies those rules to AI: which data models and agents may see, what may leave in a prompt, what must be removed from a response and how each use is recorded. The difference is speed and scale. An agent can read thousands of records in a minute, so the rules have to be enforced on every call rather than reviewed after the fact.

How is DLP for AI agents different from traditional DLP?

Traditional DLP watches email, endpoints and network egress for known patterns and usually blocks or allows. DLP for AI agents sits on the call between an agent and the tools and models it reaches. It knows which agent is acting and for which person, so the same data can be redacted for one team and allowed for another, and it can redact and let the work go on where traditional DLP would only block.

How do you redact PII from LLM prompts?

Route model traffic through an LLM gateway that inspects the prompt before it leaves. LLM Authority finds customer names, phone numbers and account details and replaces them with placeholders, so the model receives the same request without the personal data. You can strip the details and let the work go on, or stop the prompt when it shouldn't leave at all, with stricter rules for teams such as finance and support.

Can you redact PHI from MCP tool responses?

Yes. A redaction policy picks the responses it applies to, by MCP server, tool, person, group and agent, then filters them before they reach the agent. Pattern filters replace exact formats such as dates of birth and Medicare Beneficiary Identifiers with labels like REDACTED:dob, and an AI judge removes health details written as free text, such as diagnoses and medications.

Does redaction only work on fixed formats like card and Social Security numbers?

No. Pattern filters handle exact, known formats such as card numbers, routing numbers and Social Security numbers. For what a pattern can't describe, such as a diagnosis in a support note or a salary mentioned in a chat thread, an AI judge follows a plain-language instruction. Redaction works in structured fields and free text alike, and you choose what happens if the judge fails, for example blocking the response.

Can different teams have different redaction rules?

Yes. Rules are set per person, agent and team, using the groups in your identity provider. Support can see a patient's name and visit time with PHI removed, clinical staff can be exempt, and contractors can be kept off the server entirely. One rule can also name a specific agent, such as Claude Code, so the same person gets different data through different agents.

Does SecureAuth work with the DLP and AI guardrails we already run?

Yes. Agent Authority works with the data redaction and DLP services you already run, alongside its built-in filters, and with third-party AI guardrails, so their findings sit in the same decision as your policy. Content checks can also run on a model endpoint you already approve.

How do you stop AI coding agents from reading secrets?

A credential filter on your code-hosting MCP servers replaces AWS access keys and GitHub and GitLab tokens in file and code-search responses with redaction labels before the agent sees them. AI coding agent security covers the rest.

Is redacted data stored in the audit log?

No. Every call is written to Agent Trail with the agent, the person it acted for, the tool or model, the decision and the rule that made it. Redactions are recorded as counts only, so the audit trail doesn't become a second copy of the data you were protecting. You can search it or send it to your SIEM.

Does Agent Authority make us compliant with HIPAA or GDPR?

No product makes an organization compliant on its own. Rules such as HIPAA, GDPR and PCI DSS are why many teams need to control what AI agents and models see. Agent Authority gives you the technical controls those programs ask about: redaction of personal, health and payment data, access decided per person and agent and a record of every call. Your compliance team decides how they fit your program.

Keep sensitive data out of every agent call

14-day free trial

Your own Agent Authority tenant. Run it in the SecureAuth-hosted cloud, a dedicated or private cloud, self-hosted or air-gapped.

Updated