Together AI: AI Coding Agent Accidentally Pushed a Key to GitHub, and an Outsider Downloaded Prompts
A coding agent accidentally committed a key to Together AI's private storage on Hugging Face to a public GitHub repository, and an outsider used it to download customer test prompts from June and July 2026. It's like leaving a safe key in an open mailbox: a scanner spotted it and the commit was rolled back, but the window was long enough. If you send sensitive data through the API, enable ZDR, a mode in which requests are not stored, and don't put passwords into prompts.
AI-processed from Together AI Blog; edited by Hamidun News
Imagine leaving your safe key in an open mailbox. That is essentially what an AI coding agent at Together AI, a platform that runs open models in the cloud, did, only in digital form. The agent pushed a service access key for a private Hugging Face repository to a public GitHub repository (this is called "committing"). The private repository held sample customer prompts from June and July 2026. While the key was exposed, an outsider managed to use it and download those samples. For anyone who lets coding agents into their repositories and sends real data to external APIs, this is nearly a textbook case.
What happened
It started with routine work. To test its software, Together AI collected sample prompts that it was permitted to use for that purpose and stored them in a private repository on Hugging Face. The samples came from requests to the MiniMax M3, GLM-5.2 and Kimi K2.6 models on its serverless service. This is the term for a mode in which customers simply send requests through the API and the company manages the infrastructure. The main request dates were June 24 and July 4, 8 and 14, 2026.
Then, while developing the public GitHub repository xorl, a coding agent committed credentials that granted access to that same private repository. An automated scanner found the key, and the commit was rolled back. But while the key was accessible, an outsider used it and downloaded the stored samples.
Next, the intruder followed a familiar playbook: he combed the data for passwords and keys to other services and checked whether they worked. Five of the accounts he tested turned out to be canaries, decoys from a security experiment run by Dragos. Such keys open nothing of value, but they signal when someone uses them.
What is known about the scale
Together AI brought in an outside forensics firm and law enforcement. According to the company, no leaks of prompts other than this one were found. It sees no grounds to believe the attacker passed the data on, and it understands that the data has already been deleted.
There is an important caveat: the review of the contents of the affected prompts is ongoing. It is not yet known whether the samples contained personal data or other customer information. The company promises to notify anyone affected directly. It has already apologized.
What the company changed
Together AI cites four measures:
- keys with limited permissions or short lifetimes, and where possible, access without keys at all;
- checks of all pull requests for secrets;
- ZDR (zero data retention) by default for new accounts: customer requests are not stored;
- updated rules for working with coding agents.
The last point is the most telling. An agent given access to a repository cannot tell a configuration file from a secret unless it is warned and protected. The scanner did work here, but the exposure window was still long enough.
What developers should do
First: if you send sensitive data through the API, enable ZDR in your account settings. New accounts have it on by default, while older ones need to enable it manually. When requests are not stored, there is nothing to steal from storage.
Second: do not put real passwords, tokens or keys into prompts. The first thing an attacker does with a stolen archive of prompts is search it for secrets to other services. Anything that has ever been sent in a request may end up in someone else's logs.
Third: review the processes around coding agents. Use short-lived, narrowly scoped keys. Enable secret scanning at the commit and pull request level, not only after publication. Protect sample prompts from real users as strictly as you would a production database.
The results of the prompt content review are still pending, and it is not yet known how many customers will receive direct notifications. But the lesson for the industry is already clear: the weak link was not the models but the familiar development pipeline, accelerated by automation.
- Nov 21, 2026Eight Open-Source Tools for AI Agents: Browser Use Leads with 121,000 Stars
- Nov 21, 2026Thinking Machines Releases Open-Weight Inkling 2 Mini with 26.4B Parameters for 32 GB Macs
- Nov 21, 2026AURA 2.0 Video Model Starts Talking on Its Own: Response Audio Arrives in 344 ms
- Nov 21, 2026Anthropic raised a record $81.3B in its IPO, but the founders kept control
Key Questions Answered
What happened?
It started with routine work. To test its software, Together AI collected sample prompts that it was permitted to use for that purpose and stored them in a private repository on Hugging Face. The samples came from requests to the MiniMax M3, GLM-5.2 and Kimi K2.6 models on its serverless service. This is the term for a mode in which customers simply send requests through the API and the company manages the infrastructure. The main request dates were June 24 and July 4, 8 and 14, 2026.
What is known about the scale?
Together AI brought in an outside forensics firm and law enforcement. According to the company, no leaks of prompts other than this one were found. It sees no grounds to believe the attacker passed the data on, and it understands that the data has already been deleted. There is an important caveat: the review of the contents of the affected prompts is ongoing. It is not yet known whether the samples contained personal data or other customer information. The company promises to notify anyone affected directly.
What the company changed?
Together AI cites four measures: keys with limited permissions or short lifetimes, and where possible, access without keys at all; checks of all pull requests for secrets; ZDR (zero data retention) by default for new accounts: customer requests are not stored; updated rules for working with coding agents. The last point is the most telling. An agent given access to a repository cannot tell a configuration file from a secret unless it is warned and protected.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.