Skip to main content
Published in Product Insights

How to Get Claude Enterprise Through a Security Review

How to Get Claude Enterprise Through a Security Review
Table of contents
  1. Which Claude plan is the security review about?
  2. What goes in the security pack for the first meeting?
  3. Who answers which question?
  4. Does Anthropic train Claude on our data?
  5. How long does Anthropic keep our data?
  6. What does Anthropic's contract commit to?
  7. How should the review handle connectors and MCP servers?
    1. How Gnar builds connections a security team can review

To get Claude Enterprise through a security review, give the reviewer Anthropic's own documents before the first meeting. Then make every answer name the plan it covers and the page that says it.

Anthropic publishes written answers on training, retention, certifications and its contract terms. Questions about your own settings and your own systems are yours to decide.

This post is for a CISO at an insurer or a regulated manufacturer who has been handed Claude to review. It covers which plan is under review, what goes in the pack, who answers which question, and the parts that stay with your own team.

Enterprise AI integration connects an AI model, such as Claude, to the systems your company already runs. U.S.-based senior engineers at The Gnar Company build those connections. Gnar is a Claude Select Services Partner in Anthropic's Claude Partner Network.

Which Claude plan is the security review about?#

Start by naming the plan and the route your people will use to reach Claude. Anthropic writes its privacy articles in two sets. One covers commercial products such as Claude for Work and the Anthropic API. The other covers consumer products such as Claude Free, Pro and Max.

The two sets give different answers. A reviewer who cites the wrong set will record the wrong rule. Training is the clearest case.

Plan or route What the provider says about training Where it says it
Claude for Work and the Anthropic API "By default, we will not use your inputs or outputs from our commercial products" to train its models Anthropic Privacy Center, commercial
Claude Free, Pro and Max It may use chats and coding sessions to improve its models if, among other cases, you choose to allow it Anthropic Privacy Center, consumer
Claude through Amazon Bedrock "your content is not used to improve the base models and is not shared with any model providers" AWS Bedrock FAQ

The first row comes from Anthropic's commercial training article. Anthropic's Privacy Center describes a Claude for Work account as a Team or Enterprise plan. The second row comes from its consumer training article. That article lists four cases, and choosing to allow it is the first one. The third row comes from the Amazon Bedrock FAQ.

If some staff already use Claude Free, Pro, Max or Ultra accounts for work, the consumer articles apply to those accounts. Anthropic's Enterprise plan page lists a network control that blocks personal or non-corporate Claude instances from enterprise networks.

What goes in the security pack for the first meeting?#

The pack is Anthropic's public documents for your plan, plus your own decisions on how Claude will be set up. This is general engineering advice: send the pack a few days before the meeting. Attach each answer to the page that states it. Here is what to include:

  1. Anthropic's certifications page. For commercial products it lists a HIPAA-ready configuration (BAA available), ISO 27001:2022, ISO/IEC 42001:2023, and SOC 2 Type I & Type II.
  2. A request for the reports themselves. The same page says copies of the compliance documentation are requested through Anthropic's Trust Portal.
  3. The Commercial Terms of Service. As of November 2026 the page shows "Effective Jun 17, 2025".
  4. The Data Processing Addendum (DPA). As of November 2026 it shows "Effective Feb 24, 2025". Its Schedule 4 lists subprocessors, the entities Anthropic engages to process customer personal data.
  5. The commercial training and retention articles. These answer the training and retention questions for commercial plans.
  6. The Enterprise retention article and the plan page. These cover the settings your own Owners control.
  7. The zero data retention article. Include it if your developers will call the API or use Claude Code on your Enterprise plan.
  8. Your own decisions. List the retention period, the feedback setting, how people sign in, and which connectors are switched on.
  9. A list of every system Claude will reach. Note what it may read and what it may change in each one.

BAA is short for business associate agreement.

In Anthropic's announcement of Jan 13, 2025, it calls ISO 42001 "the first international standard outlining requirements for AI governance."

The certifications page covers commercial products such as Claude for Work and the Anthropic API. For consumer plans it points to a separate article, which is one more reason the plan goes first.

Who answers which question?#

Three parties answer a Claude security review: Anthropic, your own Claude Owners and admins, and whoever builds the connections to your systems. Questions about Anthropic's own controls go to Anthropic in writing. Questions about settings and connections stay inside your company.

Question Who answers Where the answer lives
Audit reports and certificates Anthropic Trust Portal and DPA section F
Breach notice and subprocessors Anthropic DPA sections C and G
Is our data used for training? Anthropic, then your Owners for the feedback setting Privacy Center and Data and Privacy settings
What retention period do we set? Your Primary Owner or Owners Data and Privacy settings
Who can sign in and use what? Your identity team and Claude admins SSO, SCIM and roles
Zero data retention for the API Anthropic's sales team Your agreement
What can Claude reach in our systems? Your Claude admins and whoever builds your connectors Connector access rules and logs

SSO is single sign-on. NIST describes it as one account used "to access multiple applications." SCIM is the System for Cross-domain Identity Management. RFC 7644 describes it as "an HTTP-based protocol" for "managing identities in multi-domain scenarios." Anthropic's Privacy Center says requests for zero data retention go to its Sales Team. It also says they are "reviewed and applied on a per-organization basis", so a company with several Claude organizations names each one.

Security sign-off is one of the approvals that sets a rollout's calendar, covered in our post on how long a Claude Enterprise rollout takes.

Does Anthropic train Claude on our data?#

Not by default, on Anthropic's commercial products such as Claude for Work and the API. Anthropic's commercial training article says: "By default, we will not use your inputs or outputs from our commercial products (e.g. Claude for Work, Anthropic API, Claude Gov, etc.) to train our models."

The same article names the exception. "If you explicitly report feedback or bugs to us (e.g. via our thumbs up/down feedback button), or otherwise choose to allow us to use your data, then we may use your chats and coding sessions to train our models." It says feedback stores the whole related conversation for up to 5 years.

A Primary Owner or Owner of a Team or Enterprise plan can switch member feedback off. The setting is called Rate chats, under Organization settings, then Data and Privacy. Record whether it is on or off in the pack, because the article ties the feedback case to that button.

A second route into training is opt in only. Anthropic's Development Partner Program lets an organization share Claude Code sessions. The help center says that under the program "only your Claude Code input and output tokens from the first-party Claude API are provided to us." It says that data is stored for up to two years. It also says "Accounts on our zero data retention agreement are not eligible for the program."

How long does Anthropic keep our data?#

On Claude Enterprise, data is retained indefinitely unless a custom retention period is set. An Owner can set one of at least 30 days. The Enterprise retention article lists products the period does not cover, such as standalone Claude Design. The answer changes with the product. For API users, Anthropic says it automatically deletes inputs and outputs on its backend within 30 days of receipt or generation, with listed exceptions.

Product or case What Anthropic states Source
Anthropic API Inputs and outputs deleted on the backend within 30 days, with listed exceptions Commercial retention article
Claude Enterprise Retained indefinitely unless a custom period is set, minimum 30 days Enterprise retention article
A deleted chat on Claude for Work Removed from chat history immediately, deleted from back-end storage within 30 days Commercial retention article
Content flagged for a Usage Policy violation Inputs and outputs up to 2 years, classification scores up to 7 years Commercial retention article
Feedback you submit Data from that submission kept for 5 years Commercial retention article

The commercial retention article lists the API exceptions. For Claude for Work, it says "we retain your chats and coding sessions in the product to provide you with a consistent product experience."

The Enterprise retention article says you must have a Primary Owner or Owner role to set the period. One line in it changes how the setting works: "Project retention always takes precedence over chat retention for chats inside a project." Projects are retained indefinitely by default, so set the project period too.

Zero data retention, or ZDR, is an arrangement with Anthropic. Under it, Anthropic does not store a customer's inputs or outputs "except where needed to comply with law or combat misuse or harm." Anthropic says ZDR is subject to its approval. Its API retention page names six models, Claude Fable 5.2, Mythos 5.2, Fable 5.1, Mythos 5.1, Fable 5 and Mythos 5. It says they "require 30-day data retention and are not available under ZDR unless expressly authorized by Anthropic." Anthropic's ZDR article says it applies to "eligible Anthropic APIs, Anthropic products that use your Commercial organization API key (including Claude Code accessed via the API), and Claude Code for Enterprise plans." It says that under these arrangements, Anthropic still retains User Safety classifier results to enforce its Usage Policy.

The Enterprise chat interface is not on that list. Anthropic's API and data retention page says the Claude Teams and Claude Enterprise product interfaces "are not ZDR-eligible." It names one exception, Claude Code on Claude Enterprise. That product has its own ZDR offering for qualified accounts.

What does Anthropic's contract commit to?#

Anthropic's Commercial Terms point to its Data Processing Addendum, which adds written breach notice within 48 hours and annual audits by external auditors. It also gives 15 days to object to a new subprocessor. The Terms say data submitted through the Services "will be processed in accordance with the Anthropic Data Processing Addendum."

Anthropic's processor article says a Team or Enterprise account is created under the Commercial Terms of Service. Under those terms, it says, the customer is the "Controller" of the data its users submit. It says "Anthropic acts as a 'Processor' of the data on behalf of the customer." In plain terms, your company decides how the data is used. Anthropic handles it on your instructions.

These are the DPA clauses a reviewer will look for, in the DPA's own words:

  • Breach notice (section G.1). Anthropic will notify the customer in writing "without undue delay, but in any event within 48 hours, after becoming aware of any Security Breach."
  • Outside audits (section F.1). "Anthropic is audited annually against known, established industry standards performed by external auditors." On written request, it will provide audit reports or certificates, such as a SOC 2 report, "to the extent available."
  • Your own audit (section F.2). On written request and at the customer's expense, Anthropic will permit the customer to audit its applicable controls, under the conditions listed. One condition is that a similar audit has not been conducted in the prior twelve months. That condition does not apply when there are indications of non-compliance or when the audit is required by a supervisory authority or other regulatory authority responsible for the enforcement of Applicable Data Protection Laws.
  • Subprocessors (section C.3). Anthropic will give reasonable notice before a new subprocessor gets access to customer personal data. The customer may object on reasonable data privacy or data security grounds within fifteen days of that notice. If it does not, the customer is deemed to consent to the new subprocessor.

Put the clause numbers in the review notes. When the DPA changes, a reviewer can then check the same clauses in the new version.

How should the review handle connectors and MCP servers?#

Review each connection from Claude to your own systems as its own item. Anthropic's help center says: "Custom connectors allow you to connect Claude to arbitrary services that have not been verified by Anthropic."

In Gnar's integration work, most connections are small programs called MCP servers. MCP stands for Model Context Protocol, an open standard for connecting AI to software. Each lets AI use one system, only in the ways you allow.

Anthropic's custom connector guide includes these points for a network reviewer. First, "Claude connects to your remote MCP server from Anthropic's cloud infrastructure, rather than from your local device." Second, the server must be reachable over the public internet from Anthropic's IP ranges. Third, "Malicious MCP servers may include hidden instructions that try to make Claude perform unintended actions."

Zero data retention does not reach these connections. Anthropic's API and data retention page says "Data processed by third-party websites, tools, or other integrations is not covered." So each connector needs its own answers: what it can read, what it can change, where it runs and what it logs.

How Gnar builds connections a security team can review#

At The Gnar Company, the AI reaches your ERP through an MCP server or an API with its own rules. Seeing a record and changing one are separate permissions. Any change goes through one checkpoint with its own checks, not straight into the ERP.

A log shows what Claude looked up and changed, so your security team can review it. Before launch, your security team gets the access rules and the request log to review.

The work starts in Gnar Discovery and Design. There, Gnar lists each system AI needs to reach, what it may see and change, and who signs off.

Gnar has worked with MCP connectors at a client. The Gnar Company ran Gnar Immerse training at GSC Enterprises. There, MCP connectors for Snowflake, Atlassian, Microsoft 365 and GitHub brought GSC's real systems into the loop for its engineers using Claude Code.

Frequently asked questions

Does Anthropic train Claude on Claude Enterprise data?

No, not by default. Anthropic's Privacy Center says it will not use inputs or outputs from its commercial products to train its models by default. Those include Claude for Work and the Anthropic API. The exception is data you choose to share, such as thumbs up or down feedback. Owners of a Team or Enterprise plan can turn member feedback off with the Rate chats setting.

What certifications does Anthropic list for its commercial products?

Anthropic's certifications page lists a HIPAA-ready configuration with a BAA available, ISO 27001:2022, ISO/IEC 42001:2023, and SOC 2 Type I and Type II. The page covers commercial products such as Claude for Work and the Anthropic API. It says copies of the compliance documentation are requested through Anthropic's Trust Portal.

How long does Claude Enterprise keep chats by default?

Indefinitely, unless an Owner sets a period. Anthropic's help center says Enterprise data is retained indefinitely unless a custom retention period is set. A Primary Owner or Owner can set one, with a minimum of 30 days. The article lists products the period does not cover. Chats inside a project follow the project's retention period, and projects are retained indefinitely by default.

Does zero data retention cover the Claude Enterprise chat interface?

No. Anthropic's developer docs say the Claude Teams and Claude Enterprise product interfaces are not ZDR eligible. Claude Code on Claude Enterprise is the exception, with its own ZDR offering for qualified accounts. Anthropic's Privacy Center says these arrangements are subject to Anthropic's approval.

How soon must Anthropic report a security breach?

Within 48 hours at the latest. Anthropic's Data Processing Addendum says Anthropic will notify the customer in writing without undue delay, but in any event within 48 hours, after becoming aware of any Security Breach. It also says Anthropic will reasonably cooperate with the customer's investigation.

Who should answer a security reviewer's questions about Claude?

Three parties answer them. Anthropic answers for its own controls, such as audit reports, subprocessors and breach notice. Your Claude Owners answer for settings, such as retention and feedback. Your Claude admins and whoever builds your connectors answer for what Claude can read and change in your own systems.

Does The Gnar Company build Claude connections a security team can review?

Yes. The Gnar Company connects Claude to internal systems through MCP servers and APIs that your company controls. A log shows what Claude looked up and changed, so your security team can review it. Start with Gnar Discovery and Design. Two weeks, $15,000, and you walk out with a build plan that has a fixed price on it.

Stay In The Loop

Never miss an update. Get the latest blogs, webinars, and expert insights straight to your inbox from The Gnar.