Understand this story
On November 21, 2026, Dragos Chief Research Officer HD Moore sent nearly 2 million requests containing traceable digital decoys to multiple AI service providers, embedding AWS access keys, passwords for services such as SSH, unique URLs, DNS names, and references to private Git repositories in prompts sent exclusively to individual providers. The same day, Together AI disclosed that samples of some customer prompts submitted to its serverless inference service in June and July 2026 had been downloaded by an unauthorized party; the company has strengthened controls over credentials and development processes and made zero data retention the default setting for new accounts.
AI-generated from public reports
Developments
2 confirmed developments
HD Moore Plants Digital Canaries in Prompts Sent to AI Services
Source:IT Home · AITogether AI Discloses Unauthorized Download of Customer Prompt Samples
Source:Together AI Blog
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Nov 22
Unite.AISelected
Together AI Discloses That Credentials Leaked by a Coding Agent Led to Customer Prompt Samples Being Downloaded
Together AI disclosed on November 21, 2026, that while a coding agent was developing the public xorl GitHub repository, it committed credentials for a private repository, allowing an unauthorized party to download some customer prompt samples. Stored in a private Hugging Face repository, the samples involved requests submitted to MiniMax M3, GLM-5.2, and Kimi K2.6; the downloader also searched for and tested credentials for other services found in the samples. The company has introduced short-lived or narrowly scoped credentials, credential checks for pull requests, and controls for coding-agent development workflows, and has made zero data retention the default option for new accounts.
Unite.AIRelated discussion
Unite.AI joined this discussion
This post contributes to heat; it is not a formal report or confirmed fact.
Nov 21
IT Home · Artificial IntelligenceSelected
Nearly 2 Million Decoy Requests Expose Privacy Risks in the AI Service Chain, With Leaks Possible Even Under Zero Data Retention Settings
Dragos Chief Research Officer HD Moore sent nearly 2 million requests containing traceable decoys to multiple AI service providers, uncovering issues including stolen prompts, data leaks by resellers, and failures to honor privacy settings. At least three providers showed signs of compromise, and some decoy keys were used to make tens of thousands of cloud service API calls within hours of submission. Private URLs in some requests marked for zero data retention or no data collection were still accessed days later.
IT Home · AIRelated discussion
IT Home · AI joined this discussion
This post contributes to heat; it is not a formal report or confirmed fact.
Together AI BlogFirst-partySelected
Together AI Discloses June and July 2026 Request Prompt Leak
Together AI disclosed that samples of prompts submitted to its serverless inference service in June and July 2026 were downloaded by an unauthorized party. The incident originated when an AI coding agent, while developing the public GitHub repository xorl, committed credentials that provided access to a private Hugging Face repository. The affected samples came from MiniMax M3, GLM-5.2, and Kimi K2.6 requests. Together AI has strengthened controls over credentials and development processes and made zero data retention (ZDR) the default setting for new accounts.
Together AI BlogRelated discussion
Together AI Blog joined this discussion
This post contributes to heat; it is not a formal report or confirmed fact.
Ars Technica
Researchers Planted 2 Million Canaries in the AI Supply Chain and Found Widespread Leakage of Sensitive Data
Researcher HD Moore planted 2 million digital canaries in AI service requests and found that the supply chain comprising routers, gateways, and inference services can leak highly sensitive data. The experiment separately embedded AWS access keys, passwords for services such as SSH, unique URLs, DNS names, and references to private git repositories in prompts sent to a single provider, then traced the leakage paths through controlled infrastructure.
Heat trend for this story
- Current
- 27
- Period peak
- 27 Nov 22 08:00
- 24-hour change
- —
Move the pointer or select the chart to inspect hourly heat. Use the left and right arrow keys with a keyboard.
AI Trends