Zero Data Retention

1. What zero data retention means

Under zero data retention ("ZDR"), the prompts, messages, tool results, and files you send ("Inputs") and the completions our models return ("Outputs") are processed in volatile memory only, for as long as it takes to generate and return the response, and are discarded when the request completes. They are not written to disk, databases, logs, backups, or archives. They are not used to train, fine-tune, or evaluate any model. They are not shared with another customer or with a third party.

In-memory caches used to accelerate inference hold key/value representations and content hashes for the cache TTL only. They are never written to durable storage and do not survive a restart of the serving pool.

Operational metadata is separate from content. Token counts, request counts, timestamps, latency, HTTP status codes, and model identifiers contain no Input or Output content. We retain that metadata on every tier, including ZDR, for billing, capacity planning, security, and reliability.

2. Who gets ZDR

Full ZDR is available on enterprise tiers only. We enable it per organization or per API key after a Zero Data Retention Addendum is signed. It is not a request header and not a setting you can toggle in the dashboard. Email sales@morphllm.com to start one.

A dedicated endpoint is not ZDR by itself. Dedicated serving isolates your traffic onto single-tenant infrastructure; retention still follows your tier until the addendum is signed.

3. Minimal retention on paid tiers

Free and paid traffic runs under minimal retention rather than ZDR:

• Bounded window: Input and Output content is kept for no longer than 30 days, then permanently deleted. Nothing is kept past that window except as described in Section 4.
• Purpose-limited: We use retained content only to serve your request, to let you inspect and replay your own traffic in the dashboard, to debug issues you report, and to investigate abuse. Never for training.
• Limited access: Access is restricted to authorized personnel under written confidentiality obligations, and is audit-logged.
• Flagged prompts are the exception: Prompts our safety classifiers flag as attacks on systems you do not own are retained past the 30-day window. See Section 4.

4. The safety exception: flagged prompts

We flag requests that read as an attempt to compromise a third party — unauthorized access against services you do not own, credential harvesting, or malware aimed at someone else's infrastructure. Those Inputs and the corresponding Outputs are retained for safety review, abuse investigation, and response to legal process.

This exception applies on every tier, including enterprise ZDR. It is bounded:

• Retention is limited to the specific flagged requests, not your account's wider traffic.
• Flagged content is never used to train, fine-tune, or evaluate a model.
• It is deleted once the review closes, unless law or valid legal process requires us to keep it.
• It stays subject to the confidentiality and security obligations of your agreement for as long as we hold it.

Because ZDR traffic leaves us no content in the ordinary course, we do not run routine content-based moderation over it. Ordinary abuse controls on ZDR accounts rely on operational metadata, and this exception is what lets us act when a specific request is flagged.

5. Privileged-access models

morph-dsv41flash-abliterated (DeepSeek V4.1 Flash Abliterated) is DeepSeek V4.1 Flash with its refusal directions ablated. It is gated, and access carries its own monitoring terms.

• Approval required: Requests from accounts without access return 403. Access is requested at morphllm.com/privileged-access, reviewed by a person, and granted to named individuals at a verified organization.
• Attested at request time: Every applicant confirms they are authorized to request access for their organization, that the model will be used only for the use case they described and in compliance with our Terms of Service, and that requests and usage may be logged and reviewed for safety and that Morph can revoke access at any time.
• Retention follows your tier, and the safety exception always applies: Paid accounts run under Section 3, enterprise accounts under their addendum. On this model specifically, the Section 4 flagging and retention path is a condition of access on every tier — including ZDR accounts — and we will not waive it.
• Revocation: We revoke access when traffic conflicts with the stated use case or the attestations above, without notice.

The intended use cases are defensive: security research, red teaming systems you are authorized to test, AI safety and alignment work, evaluation and benchmarking, trust and safety, and interpretability. The model is not for attacking third parties, and the flagging path exists to catch traffic that is.

6. Other permitted processing

Beyond Section 4, we may retain or disclose Inputs and Outputs only to the extent needed to:

• comply with applicable law, regulation, or valid legal process, after notifying you where we are legally permitted to;
• investigate fraud or a violation of our Terms of Service or Acceptable Use Policy, limited to the minimum data the investigation requires; or
• protect the safety, security, or integrity of the services, Morph, or a third party.

Content retained under any of these is never used for model training without your separate written consent.

7. Verification

Because ZDR accounts leave no stored content, no deletion step is needed to remove it. On written request (up to once per quarter), we will confirm in writing that your account is configured for ZDR and that no Input or Output content persists in our systems.

Paid-tier accounts can request early deletion of retained content by emailing privacy@morphllm.com. Deletion removes your dashboard history and limits our ability to debug issues you report.

8. Contact

ZDR addendum and enterprise terms: sales@morphllm.com
Privacy inquiries and deletion requests: privacy@morphllm.com

Address:
AutoInfra, Inc. d/b/a Morph
San Francisco, CA 94109
United States