IPCRIME Breach Response Center | November 22, 2026. You used a model provider or AI gateway, and it now reports that prompts may have been copied, retained unexpectedly or forwarded to another service. If you placed live credentials or confidential information in those prompts, act on the information that was actually sent. A provider warning is not proof that every request was stolen.
1. Establish whether your request used the affected route
Check your application logs, invoices and provider account notices for model name, provider, date, route and retention mode. Do not rely only on a remembered chatbot name; a gateway can use different downstream hosts for different requests. Read the OpenRouter announcement and, where relevant, the separate vendor notices.
2. Compare precise disclosure windows
Phala advises treating DeepSeek V4 Flash 0731 requests from October 21 00:00 UTC to November 8 05:35 UTC as potentially exposed. Novita warns about MiniMax requests between May 28 and November 11. Together AI is reviewing prompt samples from June and July. Those are different events, and their timelines should not be merged into a single universal incident.
3. Revoke secrets that may have appeared
- Replace API keys, cloud access tokens, database passwords or other active secrets included in the potentially affected prompt. Use the credential issuer’s official management interface.
- Prefer a staged rotation where critical services would otherwise break, and review usage logs for unexpected access before and after rotation.
- Revoke stale tokens and sessions, review privilege scopes and update the systems that legitimately use the replaced credentials.
- Do not post any real keys or prompt samples publicly to get help verifying exposure.
4. Assess sensitive content separately
Determine whether personal data, confidential source code, customer documents or private investigative material were included. This evaluation can inform privacy notifications or contract obligations. A prompt containing a public code example is different from one containing a real customer record.
5. Prevent repeat exposure
Limit AI context to necessary data, use synthetic samples for testing, require retention and processor review before sensitive work, and favor short-lived read-only credentials. Test fallback procedures so production systems are not bound to a provider whose handling you cannot verify.
Sources and scope
OpenRouter notice; Phala incident notice; Novita notice; Together AI notice. This is practical incident-response guidance, not a claim of direct forensic testing.