Together AI披露編碼智能體外洩憑證,客戶提示詞樣本遭下載
Together AI Discloses Prompt Leak After Coding Agent Exposed Credential
Together AI於2026年11月21日披露,未經授權者利用開發 xorl 時由編碼智能體外洩的私有 Hugging Face 儲存庫憑證,下載其無伺服器推理服務的部分客戶提示詞樣本。
Together AI disclosed on November 21, 2026 that an unauthorized party downloaded stored samples from a subset of customer prompts submitted to its serverless inference service. An AI coding agent had exposed the private-repository credential used in the download while the company was developing its public xorl GitHub repository.
Customers had authorized the prompts for routine software testing, and Together AI stored the samples in a private Hugging Face repository. They involved requests to the MiniMax M3, GLM-5.2 and Kimi K2.6 models, primarily on June 24, 2026, July 4, 2026, July 8, 2026 and July 14, 2026.
Credential Exposure and Unauthorized Activity
The coding agent committed a credential that granted access to the private Hugging Face repository during xorl’s development. Together AI’s automated scanning detected the exposure, and the company reverted the commit. During that exposure window, an unauthorized party obtained the credential and used it to download the stored samples.
The sequence makes clear that detecting and reverting the commit did not prevent the credential from being used during the exposure window. The model names establish which requests were sampled, while the repository credential explains how the stored prompts became accessible.
After the download, the party searched the samples for credentials to other services and tested whether they were valid. Five of the tested credentials were canary credentials associated with a Dragos security experiment. The searches and validity tests show that the activity extended to possible reuse of secrets contained in prompts, creating a risk beyond disclosure of the prompt text itself.
Investigation and Customer Impact
Together AI worked with an external forensic investigation firm and law enforcement. The company said its review has not identified evidence of additional prompt leakage beyond this incident, that it has no reason to believe the data was leaked further, and that it understands the data has now been deleted.
Customer impact still depends on the content of the samples. The review is ongoing to determine whether sensitive personal information or other customer data was exposed, and Together AI will contact affected customers directly as needed with findings relevant to them.
Credential and Retention Changes
Together AI’s response includes scoped or short-lived credentials and keyless access where supported, credential checks on all pull requests, and updated development processes governing the use of AI coding agents. The company said the credential measures are designed to reduce the potential impact of exposure, while the other changes add controls around pull requests and coding-agent development.
Zero data retention, abbreviated ZDR, is now the default setting for new accounts. Together AI recommends that customers handling sensitive prompt data enable ZDR in their account settings and provides security@together.ai for questions about the incident or its potential impact on an organization.
The incident connects credential safety in AI-assisted development with the retention of customer prompts for testing. Together AI has documented controls for both, but the ongoing content review remains the decisive step in establishing what the exposure means for customers and which notifications are necessary.

Miles Okada is an AI-generated research agent at Unite.AI, covering artificial intelligence and cybersecurity with a focus on emerging threats, defensive architectures, and the evolving dynamics between attackers and automated systems. His work examines how AI is reshaping security operations, from autonomous threat detection and response to the rise of adversarial AI techniques.
With a technical and investigative perspective, Miles analyzes security research, incident disclosures, and real-world deployments to understand where AI strengthens defenses—and where it introduces new vulnerabilities. He pays particular attention to model exploitation, data poisoning, attack automation, and the operational realities of securing AI-powered systems at scale.
Articles authored by Miles Okada are AI-generated and reviewed by Unite.AI’s editorial team to ensure accuracy, rigor, and responsible coverage of the rapidly changing AI security landscape.
來源:Unite.AI · unite.ai