Security event: Prompt leakage for requests in June and July of 2026
Source
together.ai
Date
Why it matters
Prompts sent to Together AI's serverless inferenceModel inference that scales on demand without the developer reserving or managing a continuously running server.Full definition → in June and July 2026 may have been exposed after a coding agent committed a repository credential. Engineers should check their exposure and scan agent commits for secrets.
Key takeaways · AI-distilled
Together AI says an AI Agent modeA mode in which one agent can inspect, edit, run commands, and iterate toward an outcome.Full definition → working on its public GitHub repository, xorl, committed a CredentialSomething presented with a request to prove that the sender is a particular identity, such as a password or a generated key.Full definition → for a private Hugging Face repository that stored PromptInformation or instructions supplied to a model; in a chat, the request you type is often called your prompt.Full definition → samples kept for routine software testing.
The exposed samples came from serverless requests to MiniMax M3, GLM-5.2 and Kimi K2.6, mainly on June 24, July 4, July 8 and July 14, 2026. An unauthorized party used the leaked credential to download them during the exposure window.
After downloading the prompts, the party searched them for credentials to other services and tested whether those worked. Together AI says five of the tested credentials were canaries tied to a Dragos security experiment.
Together AI's fixes: scoped or short-lived credentials (keyless where supported), credential checks on every pull request, zero data retention by default for new accounts, and updated rules for AI coding agents.
Review of the downloaded prompts is ongoing, so Together AI has not yet said whether sensitive personal data was exposed. It reports no evidence of other prompt leakage and understands the downloaded data has been deleted.
Terms in this piece · Glossary
Prompt — Information or instructions supplied to a model; in a chat, the request you type is often called your prompt.
Agent mode — A mode in which one agent can inspect, edit, run commands, and iterate toward an outcome.
Credential — Something presented with a request to prove that the sender is a particular identity, such as a password or a generated key.
serverless inference — Model inference that scales on demand without the developer reserving or managing a continuously running server.