Together AI: verifying prompt retention and credential access
Together AI’s disclosure brings prompt samples and development credentials into the same risk assessment. Learn how to check permissions, retention and closure evidence for each account.
Читати португальськоюBy Pentesty
Define the data before assessing impact
In its November 21 disclosure, Together AI confirmed unauthorized access to retained prompt samples after a coding agent exposed a credential in a public repository. The request dates cited by the company do not establish when the intrusion occurred. Its account of the response measures is a vendor statement.
To assess your company’s environment, start with the data path: application, contracted account, inference destination and copies used for diagnostics. Record which projects send customer information, code or internal documents. A disabled integration may have left historical records; an active integration may use an account with settings that differ from the others.
The public notice guides the investigation, while decisions about your environment require specific evidence. Ask the provider to identify the accounts and periods covered by its assessment, and record what remains unconfirmed. Avoid turning an incomplete response into a statement that your business was unaffected.
Verify retention for each account and destination
Together AI reported zero data retention as the default for new accounts. That statement does not establish existing account settings or deletion of historical copies.
Check the policy that applies to each account, project and destination, including diagnostic exceptions and contracts. Record the observed setting, the review date and who can change it. If the interface does not expose a processing stage, keep that stage open and request a response defining the data and period it covers.
To check your own integration, send only a synthetic marker without personal data or secrets, in an authorized environment. Follow the records under your control and document which copies you can observe. A marker’s absence from a local dashboard does not prove its absence from the provider’s systems.
Repeat the review when accounts, models or integrations change. A setting checked during a pilot may not represent the production account. The Fireworks and Notion case provides a relevant comparison of provider boundaries, involving a separate event and scope.
Document the decision and outstanding work
Assign an owner to collect account settings, team tests, provider responses and open exceptions. Each piece of evidence should identify what was directly observed and what depends on an external statement. This distinction makes it possible to revisit the conclusion when additional information emerges.
Before closing the assessment, compare the potentially affected dataset with the application’s actual use. If periods, accounts or destinations are missing, assign an owner and deadline to the gap. Keep sufficient records to justify the decision without copying sensitive prompts into the investigation ticket.
Include the transition from the development environment to the external service in the next exercise. Our guide to repositories and build execution complements that work. The expected outcome is a traceable decision with explicit limitations and actions that can be verified.
OpenRouter: check routing and zero-retention eligibility
Editorial update: November 22, 2026
In its November 22 disclosure, OpenRouter said it had removed three providers from its platform and withdrawn 125 endpoints from Zero Data Retention (ZDR) eligibility. The ZDR removals include endpoints removed as a precaution. That number does not mean 125 confirmed leaks. The company said the findings showed no breach or unauthorized access to OpenRouter itself and that it had notified potentially affected customers.
Technical verification of network paths remains underway. OpenRouter requires assurances and validation before restoring eligibility to removed endpoints. The provider announcements linked on its page do not identify which three providers were removed.
To apply this information to your integration, record the provider actually used, its endpoint, the observed ZDR eligibility and permitted fallback routes. In an authorized test environment, use synthetic data and simulate unavailability to check whether switching destinations respects those restrictions. Retain the timestamp, configuration, result and owner without including sensitive prompts.
This exercise checks the integration’s observable behavior. Internal processing at third parties still requires provider evidence. If the actual destination is absent from available records, keep that gap open and assign an owner to investigate.
Frequently asked questions
Does reverting a commit invalidate a credential?
A revert undoes the commit’s effect on versioned files. The team needs to check authorization at the issuer and retain evidence of revocation where applicable.
Does the new-account default establish my account’s retention?
Check the settings and contract for the account in use. A default for new accounts does not establish the state of existing accounts or historical copies.
How can we test without exposing real data?
Use an authorized environment and a synthetic marker containing no secrets. Record only observable stages and keep external gaps open as outstanding work.
What evidence supports closing the assessment?
Collect the account scope, provider responses, local tests and exceptions with assigned owners. Record the limits of each piece of evidence.
Sources and references
- Unite.AI: reporting consulted on the Together AI disclosure
- Together AI: security disclosure dated November 21, 2026
Original Pentesty analysis, published November 21, 2026. The proposed scenarios are editorial guidance, with scope and validation to be defined for each environment.