AI provider prompt theft illustrated as a hidden data diversion inside cloud inference infrastructure.

AI Provider Prompt Theft: What Zero Retention Cannot Protect

Key takeaways

  • According to The Hacker News, OpenRouter removed three providers after researchers detected prompt theft inside inference providers.
  • Some stolen requests carried Zero Data Retention flags, highlighting the difference between retention promises and protection against an active compromise.
  • The investigation identified separate risks involving compromised infrastructure, marketplace logging, and undisclosed downstream routing.
  • Organizations should verify the full processing chain, minimize sensitive prompt content, and demand evidence supporting privacy assurances.

AI provider prompt theft can expose sensitive requests inside the infrastructure that processes them—even when those requests carry a Zero Data Retention flag. According to The Hacker News, research by Dragos prompted OpenRouter to remove three providers and withdraw zero-retention eligibility from 125 endpoints, while OpenRouter said its own systems had not been breached.

The November 2026 report on inference-provider prompt theft describes a trust problem that extends beyond one gateway: customers may know which model they selected without knowing every company or system handling the request.

For organizations using hosted AI, the practical question is not simply whether a vendor promises to delete prompts. It is whether sensitive information can be intercepted, logged, forwarded, or accessed before that promise has any effect.

What happened in the AI provider prompt theft investigation?

According to The Hacker News, Dragos researchers seeded about two million prompts with unique links and credentials between July and November. They then watched for activity involving those planted secrets. The earliest of the three provider breaches began in mid-July, and none of the three providers knew about the theft until researchers notified them.

The strongest evidence came from a breach involving a couple hundred planted AWS keys. Each key had appeared in exactly one request to one provider during short windows in late October. Researchers reported that every request sent during those windows was stolen.

Some stolen keys came from requests rejected before a model ran. That detail matters: it suggests interception at the initial arrival point, rather than extraction only from model execution or stored conversation histories.

Attackers tested some keys within hours. Their activity included listing S3 buckets, searching Secrets Manager and Parameter Store, and attempting to create Bedrock API keys and additional IAM users. The report describes tens of thousands of AWS API calls.

OpenRouter said it removed three providers, notified potentially affected customers, and removed 125 endpoints from Zero Data Retention eligibility. Those 125 endpoints should not be read as 125 confirmed compromises: the review included precautionary removals. The gateway said restoring eligibility would require assurances and validation.

Why doesn’t Zero Data Retention prevent every leak?

Zero Data Retention addresses a data-handling promise. It is not, by itself, proof that nobody can access a request while it is being received or processed.

In this investigation, some stolen requests carried zero-retention flags. If an attacker can copy incoming traffic before inference begins, deleting a provider’s legitimate copy afterward does not retrieve the attacker’s copy. Encryption in transit also cannot, by itself, resolve a compromise at the receiving system where the request becomes readable for processing.

The reporting also describes concerns beyond external intrusions:

Reported exposure pattern What researchers observed What buyers should verify
Compromised inference infrastructure Planted secrets from incoming requests were used by attackers Security controls around request ingress and processing
Marketplace seller logging Low-priced sellers enabled logging and collected requests Seller identity, logging settings, and contractual accountability
Undisclosed routing A router forwarded some traffic to a marketplace without telling customers Actual downstream processors and routing restrictions
Continued prompt access Secret links were accessed well after requests were sent Retention scope, downstream copies, and deletion verification

The lesson is not that every zero-retention service is unsafe. It is that retention, access control, and routing transparency are separate questions. A procurement review that checks only the first leaves important gaps.

How did researchers distinguish leakage from ordinary processing?

A canary token is a deliberately planted marker whose later use can reveal access to the material containing it. Here, researchers placed unique links and credentials inside prompts, giving them a way to connect later activity to particular requests.

A link fetch alone needs interpretation: some AI workflows legitimately retrieve web content. Credential use, repository-cloning attempts, and repeated access long after a request provide additional context. The reported AWS activity is especially significant because it went beyond reading a prompt and into attempts to use its contents.

The Hacker News also reported that an unnamed Chinese AI lab fetched secret links from more than half of seeded prompts sent to its models between July and October. That included about two-thirds of requests marked no data collection or Zero Data Retention. Requests sent through U.S. hosts reselling the lab’s models triggered the same scraper.

In another observation, a secret link appeared in a Feishu chat 24 days after the original prompt. Researchers said they had disclosed only the first eight characters of that provider’s link, supporting their conclusion that someone had access to a cached copy of the full prompt.

These observations should not all be collapsed into one explanation. Infrastructure intrusion, seller logging, automated retrieval, and human access are different behaviors with different remediation needs. Nor does the investigation establish that every provider or every customer request was affected.

This is also distinct from prompt injection, where malicious instructions attempt to redirect model behavior. Our guide to PromptShieldBench and prompt-injection defenses covers that separate problem. A model can resist hostile instructions while its hosting infrastructure still exposes the conversation.

What should organizations verify before using hosted models?

Treat an inference service as a data-processing supply chain, not just a model name in an API response. The following checks are practical recommendations based on the reported failure modes, not additional findings from the investigation.

1. Identify every party that can receive a request

Ask the gateway to identify the inference operator, marketplace participants, subprocessors, and any downstream model service. Establish whether routing can change automatically and whether fallback routes preserve your required privacy controls.

Request a contractual restriction against undisclosed forwarding. If a provider cannot explain who handles a request, avoid sending data that requires a clearly bounded processing chain.

2. Define what zero retention actually covers

Ask whether the commitment includes request bodies, responses, error logs, debugging traces, abuse-review systems, caches, and downstream processors. Verify how rejected requests are handled—the investigation suggests those can be exposed before model execution.

Ask what evidence supports the promise and what happens when an endpoint loses eligibility. A label in a dashboard is less useful than a documented scope, enforceable commitments, and a validation process.

3. Minimize sensitive material before transmission

Remove live credentials, unnecessary personal information, and confidential documents that the task does not require. Prefer a redacted excerpt over a complete customer record, and a placeholder over a production secret.

Review reusable prompt templates, too. Structured prompt workflows, such as those discussed in our POML prompt-engineering explainer, are an opportunity to make sensitive-data checks consistent rather than leaving them to individual users.

4. Test controls without creating new exposure

Where authorized, consider synthetic canary content to test whether secrets or links appear outside their intended path. Do not put working production credentials into prompts as an improvised security test.

Ask vendors how they validate routing and retention continuously. OpenRouter said it was adapting the researchers’ approach for ongoing checks and verifying network paths to identify third-party forwarding.

For broader review criteria, consult the official resources from NIST, CISA, and OWASP. Use those resources to structure questions; do not treat any framework reference as evidence that a particular endpoint meets your requirements.

What should teams do if sensitive prompts may be exposed?

Start by determining what was sent, through which provider and endpoint, and during which period. Preserve relevant request metadata and provider notices without creating unnecessary new copies of sensitive prompt content.

If prompts contained live secrets, prioritize revoking or rotating them and reviewing associated access logs. The reported attempts to create IAM users illustrate why replacing an exposed key may not be enough: responders should also check for additional access created with it.

For confidential business or personal information, involve security, privacy, and legal teams to assess scope and applicable obligations. A provider’s lack of a public announcement should not substitute for a direct answer about your account; OpenRouter said not all affected providers had disclosed publicly.

Separate three workstreams: securing credentials, investigating data exposure, and deciding whether the processing route remains acceptable. Our coverage of AI providers disrupting malicious model abuse addresses a related but different issue: stopping abusive use of models does not establish that customer prompts are protected inside hosting infrastructure.

The Bottom Line

AI provider prompt theft exposes the limits of evaluating hosted AI through model quality, price, and a retention checkbox alone. The reported incidents show why organizations need visibility into the systems receiving requests, the companies handling them, and the evidence behind privacy commitments.

Before your next sensitive deployment, choose one production AI workflow and map its full processing route. Confirm its retention terms, remove unnecessary secrets, and ask how the provider detects unauthorized access while requests are in flight—not just what it promises to delete afterward.

Frequently Asked Questions

What is AI provider prompt theft?

AI provider prompt theft is unauthorized access to or copying of requests sent to a hosted AI service. Exposure can occur within the infrastructure receiving or processing the request, rather than through the model's answer.

Does Zero Data Retention prevent attackers from reading prompts?

Not by itself. A retention commitment does not prevent an attacker who has compromised a receiving system from copying a request before deletion or before model execution.

Was OpenRouter itself breached?

According to The Hacker News, OpenRouter said the findings showed no breach of its own systems. It removed three providers and withdrew Zero Data Retention eligibility from 125 endpoints, including some removed as a precaution.

Is prompt theft the same as prompt injection?

No. Prompt theft concerns unauthorized access to request content, while prompt injection uses malicious instructions to influence model behavior. Defenses against one do not automatically protect against the other.

What should a business ask an AI inference provider about privacy?

Ask who receives requests, whether routing introduces downstream processors, and what the retention commitment covers. Request evidence of validation, clear incident-notification terms, and an explanation of how rejected requests and diagnostic logs are handled.

What should I do if I sent an API key in an exposed prompt?

Revoke or rotate the key and review its activity for unauthorized use. Also investigate whether it was used to create additional credentials, accounts, or other persistent access.

Sources

Enjoyed this? Subscribe to the InnoVirtuoso newsletter for practical AI, security and tech insights — and share your thoughts in the comments.

Browse InnoVirtuoso for more!