deniz.in

Markets

Weather

Loading weather

· via Hacker News – Front Page (hnrss.org)

AI coding agent leaked Together AI credential, exposing stored customer prompts

Together AI says an unauthorized party downloaded stored customer prompts after an AI coding agent committed a credential for its private Hugging Face repository to a public GitHub repo.

AI coding agent leaked Together AI credential, exposing stored customer prompts

What happened

Together AI has disclosed a security incident affecting prompts sent to its serverless inference service. According to a company blog post signed by CEO and co-founder Vipul Ved Prakash, an unauthorized party downloaded stored prompt samples that Together AI had collected as part of routine software testing and kept in a private Hugging Face repository. The company characterized the collection of those samples as authorized for that testing purpose.

The samples came from requests routed to the MiniMax M3, GLM-5.2 and Kimi K2.6 models, with affected requests falling primarily on June 24, July 4, July 8 and July 14, 2026. Together AI says its review of the sample contents is ongoing, to determine whether sensitive personal information or other customer data was included, and that it will contact affected customers directly as findings emerge.

How the credential leaked

The exposure traces back to Together AI's public GitHub repository, named xorl. While that repository was being developed, an AI coding agent committed a credential that granted access to the private Hugging Face repository holding the prompt samples.

Together AI's automated scanning detected the exposed credential and the commit was reverted, but the credential was live long enough that an outside party obtained it during the exposure window and used it to download the stored samples.

What the intruder did next

According to the disclosure, the unauthorized party did not simply archive the data. It searched the downloaded prompts for credentials to other services and tested whether those credentials were valid. Five of the credentials it tried were canary credentials — decoy secrets deliberately planted to detect exactly this kind of misuse — associated with a security experiment run by Dragos.

Together AI said it has no reason to believe the data has been distributed further and that it understands the party has since deleted it. The company brought in an external forensic investigation firm and worked with law enforcement, and reports finding no evidence of prompt leakage beyond this incident.

Changes Together AI has made

In response, the company says it has tightened controls across its development workflows, credentials and customer data handling. Specific measures include:

  • Adopting scoped or short-lived credentials, and keyless access where supported, to limit the damage a future exposed credential could cause.
  • Checking every pull request for exposed credentials.
  • Making zero data retention (ZDR) the default setting for new accounts, so prompts are not stored unless customers opt in.
  • Updating the development processes that govern how AI coding agents are used.

For customers, Together AI recommends enabling ZDR in account settings if they handle sensitive prompt data, and directs questions about the incident or its potential impact to its security team.

Why it matters

Two things make this incident notable. First, the leaked credential was committed by an AI coding agent — a class of tool now embedded in many engineering teams — pushing a live secret into a public repository. Together AI's decision to rewrite its development process around such agents is an early concrete example of a vendor treating agentic coding tools as a security surface in their own right, not just a productivity feature.

Second, it underlines that prompts are customer data. Users of inference APIs routinely paste source code, internal documents and even credentials into model requests, and retained prompt logs become a concentrated target as a result. The intruder in this case explicitly mined the stolen samples for credentials to other services and checked whether they worked. Together AI making ZDR the default for new accounts effectively concedes that storing prompts is a liability; existing customers who handle sensitive material would be well advised to verify what their retention settings actually are.

  • #security
  • #together-ai
  • #ai-agents
  • #credentials
  • #llm

Related posts