---
type: "post"
id: "security-event-prompt-leakage-for-requests-in-june-74b417"
slug: "security-event-prompt-leakage-for-requests-in-june-74b417"
title: "Security event: Prompt leakage for requests in June and July of 2026"
description: "Together AI says an unauthorized party downloaded stored prompt samples submitted to its serverless inference service in June and July 2026, after a credential was exposed in a public code repository."
retrieval_nugget: "Together AI says an unauthorized party downloaded stored prompt samples submitted to its serverless inference service in June and July 2026, after a credential was exposed in a public code repository. The company is still reviewing the affected data; teams using the service should assess whether sensitive prompts were retained and enable zero data retention where appropriate."
published_at: "2026-11-22"
updated_at: "2026-11-22"
record_date: "2026-11-22"
date_kind: "published_at"
topics: ["security"]
entities: ["Together AI","MiniMax M3","GLM-5.2","Kimi K2.6","Hugging Face","Dragos","xorl","Vipul Ved Prakash"]
related_patterns: ["agent-security-moves-to-runtime-boundaries"]
novelty_label: "new_fact"
what_changed: "Together AI disclosed an incident in which an unauthorized party downloaded stored samples of some June–July 2026 inference prompts."
prior_refs: []
source_url: "https://www.together.ai/blog/security-event-prompt-leakage-for-requests-in-june-and-july-of-2026"
source_urls: ["https://www.together.ai/blog/security-event-prompt-leakage-for-requests-in-june-and-july-of-2026"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:security-event-prompt-leakage-for-requests-in-june-74b417"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/security-event-prompt-leakage-for-requests-in-june-74b417/","markdown":"https://newruntime.com/posts/security-event-prompt-leakage-for-requests-in-june-74b417.md","json":"https://newruntime.com/posts/security-event-prompt-leakage-for-requests-in-june-74b417.json"}
markdown_url: "https://newruntime.com/posts/security-event-prompt-leakage-for-requests-in-june-74b417.md"
---

# Security event: Prompt leakage for requests in June and July of 2026

## Retrieval answer

Together AI says an unauthorized party downloaded stored prompt samples submitted to its serverless inference service in June and July 2026, after a credential was exposed in a public code repository. The company is still reviewing the affected data; teams using the service should assess whether sensitive prompts were retained and enable zero data retention where appropriate.

## Summary
Together AI says an unauthorized party downloaded stored prompt samples submitted to its serverless inference service in June and July 2026, after a credential was exposed in a public code repository. The company is still reviewing the affected data; teams using the service should assess whether sensitive prompts were retained and enable zero data retention where appropriate.

## What happened
Together AI’s announcement, published November 21, 2026, describes an incident involving a subset of prompts submitted to its serverless inference service during June and July. The company says it collected samples authorized for routine software testing and stored them in a private Hugging Face repository. The samples came from requests to MiniMax M3, GLM-5.2, and Kimi K2.6, primarily on June 24, July 4, July 8, and July 14, 2026.

The announcement traces access to development of Together AI’s public GitHub repository, xorl. An AI coding agent committed a credential that could access the private repository. Together AI says automated scanning detected the exposed credential and the commit was reverted. During the exposure window, however, an unauthorized party obtained the credential and used it to download the stored prompt samples.

The company says the party searched the samples for credentials to other services and tested whether those credentials were valid. Five tested credentials were canary credentials associated with a Dragos security experiment. Together AI says it has no reason to believe the data was leaked further and understands it has now been deleted. It worked with an external forensic investigation firm and law enforcement; the company says it has not identified evidence of additional prompt leakage beyond this incident.

Together AI says its review of the affected prompt contents is ongoing, including to determine whether sensitive personal information or other customer data was exposed. It plans to contact affected customers directly as needed. The company lists several post-incident changes: scoped or short-lived credentials and keyless access where supported; credential checks on all pull requests; zero data retention (ZDR) as the default for new accounts; and updates to development processes governing AI coding agents. Customers handling sensitive prompt data are advised to enable ZDR in account settings. The announcement does not state a price or a new product version associated with these changes.

## What the archive adds
New Runtime’s earlier coverage on [an October 2026 security incident](https://newruntime.com/posts/update-on-october-2026-security-incident-c8af72/) described Fireworks AI’s disclosure that a cloud-provider breach affected account and billing information and some secrets, datasets, and incomplete models. Together AI’s announcement is a new incident, not an update to that event. The security contexts are related, but the causes and affected services differ.

## Trend signal
This incident supports New Runtime’s view that [agent security moves to runtime boundaries](https://newruntime.com/patterns/agent-security-moves-to-runtime-boundaries/). Together AI’s listed response includes credential checks, scoped or short-lived credentials, and keyless access, while also revising how its development workflows use coding agents. Those controls address access and execution conditions around the agent rather than relying on prompt instructions alone. The disclosure does not establish how effective each measure will be in practice.

## Why now
The company published its disclosure on November 21, months after the affected requests in June and July. Together AI says the review of prompt contents is ongoing and that it is determining which customers require direct notification. That means the announcement is both an incident account and an interim status update: the scope may become clearer as the review proceeds.

## Why it matters
Prompt data can be more sensitive than a team’s model choice suggests. Requests may include customer records, internal context, or secrets copied into an AI workflow; if samples are retained for testing, that creates another repository and credential boundary to defend. Together AI’s account also shows how an AI coding agent can become part of a credential exposure path, even when automated scanning detects the mistake and the commit is reverted.

For teams, the useful response is to treat retention and access as explicit design decisions. Check what prompts your provider retains and whether the account’s retention setting matches the data classification. Keep credentials out of prompts and repositories, prefer scoped or short-lived secrets, and add automated secret scanning to pull requests. These controls reduce potential exposure; they cannot establish whether data in this particular incident affected any individual customer.

## What happens next
- **Together AI will directly notify customers whose data its ongoing review finds relevant to the incident** (3 months, 75%). Watch for: customer notifications or a further company update describing the review’s findings.

## Limits and uncertainty
Together AI’s review of the affected prompt contents is ongoing, so the extent of any sensitive personal information or other customer data exposure remains undetermined. The company says it has not identified evidence of additional prompt leakage and has no reason to believe the downloaded data was leaked further, but these are statements about its investigation rather than an independent public verification of deletion. The announcement does not quantify the number of affected customers or prompt samples. It also does not describe the precise exposure window, the identity of the unauthorized party, or the results of its review of each tested credential.

## Skill to build now
This week, inventory which prompts sent to inference providers can contain personal data, customer records, or secrets. For Together AI accounts that handle sensitive prompt data, enable ZDR in account settings, as the company recommends. Separately, add a pull-request secret scan and test that it blocks committed credentials; use scoped or short-lived credentials where a workflow needs repository access.

## Reader takeaway
A prompt retained for testing is still data that can be exposed, and an agent-assisted development workflow can create the credential path to it. Verify retention settings for sensitive workloads, reduce the reach and lifetime of credentials, and do not treat a reverted secret as proof that it was never accessed.

## Open question
What did Together AI’s ongoing review ultimately find about customer data in the downloaded samples, and how many customers required direct notification?

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Security](https://newruntime.com/topics/security.md): Explore the security topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)
