Anthropic can override zero-data-retention rules

The change

On June 8, 2026, Anthropic updated its Service Specific Terms with a new effective date. Venpo detected the change the same day.

Inside the revision sits a new contractual category called Covered Models. For any model Anthropic designates as covered, inputs, outputs and related usage data may be retained for safety review. The terms state that this applies even where stricter retention arrangements exist, including reduced-retention and zero-data-retention commitments.

Anthropic's models sit inside a large share of modern SaaS products, both through direct API usage and through cloud marketplaces, so a change to its retention framework reaches far beyond its direct customers.

What changed

Two shifts stand out in the updated document:

• a new Covered Models category that allows retention of inputs, outputs and usage data for safety review, superseding stricter retention arrangements

• an exception to the assurance that marketplace integrations give Anthropic no access to customer environments on AWS Bedrock and Google Vertex

The same revision also softened the Beta Services language: beta offerings are no longer described as definitely unsuitable for production use, only as possibly unsuitable.

Covered Models can supersede zero-data-retention commitments

Zero-data-retention arrangements have been one of the main mechanisms that allowed security and legal teams to approve AI vendors. The commitment was simple: prompts and outputs are not stored, so the data exposure question largely disappears.

The new Covered Models section changes the structure of that commitment. Anthropic may designate specific models as covered, and for those models it may keep inputs, outputs and other usage-related data for safety review. The terms are explicit that this overrides stricter retention arrangements that would otherwise apply.

The stated purpose is safety review, and the mechanism is written in plain sight in the terms. The practical consequence for customers is still significant: a retention commitment that can be superseded by a vendor-side designation behaves differently from an unconditional one, and any internal approval that relied on the unconditional version is now based on an outdated document.

The cloud marketplace assurance now has an exception

The previous terms stated that Anthropic's marketplace integration does not give it access to customer environments on AWS Bedrock or Google Vertex. That statement is no longer unconditional.

For Covered Models used through a cloud platform, related data may be processed by Anthropic under the applicable data processing terms, and the customer must confirm that arrangement to the marketplace provider. Teams that chose the Bedrock or Vertex route specifically to keep model traffic inside their own cloud boundary will want to re-read this section carefully.

One layer deeper: Databricks added Anthropic as a subprocessor

Two days after the terms update, on June 10, Venpo detected a related change in a different document: Databricks added Anthropic to its subprocessor list for AI-backed services.

This means the new Anthropic terms are relevant even for companies that never signed with Anthropic directly. When a platform vendor routes AI features through Anthropic, the retention framework described above now sits one layer deeper in the stack, where it is even less likely to be reviewed.

Anthropic's own Privacy Policy was updated in the same June 8 revision cycle, including a narrower description of the account-level training opt-out and new data categories for identity verification and study participation.

Why this matters

AI vendors have moved from the edge of the stack to its center in under two years, and their legal documents are changing at a pace that traditional annual vendor reviews were never designed for.

This change is a clear example of why point-in-time approval is not enough. The clause that anchored many security reviews did not disappear; it became conditional. A team that approved Anthropic in January and re-reads the terms next January will have spent months operating on assumptions the contract no longer supports. This is exactly the silent drift covered in our analysis of the hidden risk of vendor legal changes.

Potential impact for SaaS companies

Companies using Anthropic models directly or through cloud platforms may want to review whether:

• their security and legal approval of Anthropic relied on zero-data-retention or reduced-retention commitments

• any models they use are or become designated as Covered Models, and what retention applies in that case

• their Bedrock or Vertex architecture assumptions still hold under the new marketplace exception

• platform vendors in their stack, such as Databricks, now route data to Anthropic as a subprocessor

• customer-facing privacy documentation and DPAs still accurately describe where AI data can flow

Even where the safety rationale is reasonable, the contractual structure has changed, and that belongs in a review rather than in a discovery during an audit. Structured vendor contract monitoring is what turns a change like this into an internal signal instead of a surprise.

How Venpo detected it

Venpo continuously monitors vendor legal documents and detects changes as soon as they are published. When Anthropic updated its Service Specific Terms, Venpo immediately:

• detected the new effective date and the revised sections

• identified the Covered Models category and its relationship to existing retention arrangements

• flagged the narrowed cloud marketplace assurance as a material shift

• connected the June 10 Databricks subprocessor addition to the same vendor

The full redline of the change is public on our live feed: Anthropic Service Specific Terms, June 8, 2026. Building continuous vendor risk monitoring into the compliance stack is what makes this kind of same-day detection possible.

Business outcome

Companies that caught this change early were able to:

• re-check which of their AI workloads rely on zero-data-retention commitments before routing more data through them

• ask Anthropic directly which models are designated as Covered Models under their agreements

• review their Bedrock and Vertex assumptions with their cloud architecture teams

• update vendor records to reflect Anthropic appearing as a subprocessor inside platform tools like Databricks

Instead of finding out during a customer security questionnaire or an audit, they had time to adapt on their own schedule.

Key takeaway

The most important clause in an AI vendor contract can change shape without disappearing. A zero-data-retention commitment that can be superseded by a vendor-side designation is a different commitment than the one most teams approved, and the only way to know it changed is to be watching the document when it happens. Venpo detected this change the day it was published and traced it one layer deeper into the stack two days later. A closer look at why scheduled reviews keep missing this kind of shift is in our analysis of manual vs automated vendor monitoring.