Answers from the documentation; checking which model is on duty
Hi - I answer from the OpenSmartRoute documentation: routing, the API, plans and quotas, self-hosting. Ask away, or open a support ticket if you need a person.
Grounded in the docs - follow a source before acting on it.
Together AI discloses credential leak after coding agent exposed priva - OpenSmartRoute
Together AI discloses credential leak after coding agent exposed private repo access
An unauthorized party downloaded customer prompts from Together AI’s serverless inference service after a credential was exposed. The company responded quickly and is reviewing the incident.
Key points
Credential was exposed during development of public GitHub repository.
The leak involved stored samples from models MiniMax M3, GLM-5.2, and Kimi K2.6.
The incident occurred mainly on June 24, July 4, July 8, and July 14, 2026.
Together AI now uses zero data retention as the default for new accounts.
Why it matters: Proper credential and data management reduces security risks and protects customer data in AI systems.
By OpenSmartRoute editorial · written through the router by llm-small
From Unite.AI - “Together AI Discloses Prompt Leak After Coding Agent Exposed Credential”
Announcement - Together AI reports credential exposure and data download incident
On November 21, 2026, Together AI announced that an unauthorized party downloaded stored samples from some customer prompts. The incident involved a credential leak during the development of a public GitHub repository. The company quickly responded by investigating and taking steps to limit the damage. This disclosure highlights the importance of security in AI services that handle sensitive data.
The company identified that a coding agent used in their development process exposed a private repository credential. This credential allowed access to stored customer prompts. Together AI confirmed that the incident was limited to a specific window of time. They are now reviewing the scope and impact of the leak. The company also committed to improving security measures to prevent similar incidents.
Details of the incident - How the credential was exposed and used during the window
The incident began when a commit made by a coding agent accidentally included a private repository credential. This credential granted access to a private Hugging Face repository where customer prompts were stored. The commit was part of ongoing development work on a public GitHub repository called xorl.
Despite reverting the commit, the credential remained accessible during the window when it was exposed. An unauthorized party detected the credential and used it to download stored samples. The window of exposure lasted until the company identified and reverted the commit. Reverting the commit did not immediately stop the credential from being used, as the attacker had already accessed the data.
The attacker’s activity was limited to the time when the credential was active. Once the company identified the breach, they took steps to revoke the credential and delete the stolen data. The incident underscores how quickly malicious actors can act once a credential is exposed, even if the commit is later reverted.
Customer prompt data - Types of prompts, storage, and involved models
The prompts involved in the incident were submitted to AI models for testing and development purposes. Customers had authorized these prompts for routine software testing, not for public release. The prompts were stored in a private repository on Hugging Face, a platform for sharing machine learning models and data.
Agentic AI workloads are increasing demand for on-premises hardware. Enterprises prefer owning their compute to improve security, control, and data governance.
The stored samples included requests made to several models, including MiniMax M3, GLM-5.2, and Kimi K2.6. These models are used for various AI tasks, such as coding assistance and language understanding. The prompts covered a range of testing scenarios, often involving code snippets or natural language instructions.
The storage of prompts in a private repository was intended to keep customer data secure. However, the exposure of the repository credential allowed an attacker to access these stored samples. The company is now reviewing whether any sensitive personal information or other customer data was included in the prompts. They plan to contact affected customers directly if necessary.
Unauthorized activity - Search for other credentials and testing for validity
After gaining access to the stored prompts, the attacker searched for other credentials within the data. They looked for secrets that could be used to access additional services or repositories. The attacker tested some of these credentials to see if they were still valid.
The tests revealed that five credentials were associated with a security experiment called Dragos. These credentials appeared to be canary tokens, which are used to detect unauthorized access. The activity suggests the attacker was trying to find other secrets that could be reused elsewhere.
This activity extended beyond just downloading prompts. The attacker’s search for credentials indicates an attempt to expand access to other systems. The company’s security team worked with external forensic experts and law enforcement to analyze the incident. They are still investigating whether any further data was compromised.
Investigation and response - Forensic work, data deletion, and ongoing review
Together AI launched a forensic investigation immediately after discovering the incident. They engaged an external firm to analyze how the breach occurred and what data was affected. The investigation confirmed that the stolen data was limited to the stored samples from the specific prompts.
The company stated that they have no evidence of additional prompt leaks beyond this incident. They also confirmed that the stolen data has now been deleted from the attacker’s access. The review process is ongoing to determine if any sensitive personal or customer data was exposed.
Together AI plans to notify affected customers directly if their data was involved. They are also reviewing their security policies and procedures. The goal is to prevent similar incidents in the future. The company emphasized that their response included deleting the compromised data and revoking the exposed credential.
Security measures - Credential checks, short-lived credentials, and process updates
In response to the incident, Together AI has implemented several security measures. They now perform credential checks on all pull requests to prevent accidental exposure. They also adopted the use of short-lived credentials that expire quickly after use.
The company is updating its development processes to include better controls over the use of AI coding agents. These controls aim to reduce the risk of credentials being exposed during development. They are also increasing monitoring for unusual activity that could indicate a breach.
Together AI’s measures are designed to limit the potential impact if a credential is exposed again. Using short-lived credentials means that even if a leak occurs, the window for malicious use is very narrow. The company’s focus remains on preventing future leaks and improving overall security.
Default data policies - Zero data retention now standard for new accounts
To reduce the risk of prompt data leaks, Together AI has adopted a new default policy. All new accounts now have zero data retention (ZDR) enabled by default. This policy means that customer prompts are not stored unless explicitly allowed.
The ZDR setting is meant to protect customer privacy and prevent accidental exposure of sensitive data. Customers who handle sensitive prompt data are encouraged to enable ZDR in their account settings. The company provides a dedicated email for questions about data security and incident impact.
This policy change reflects a broader effort to improve data privacy and security. It aims to make prompt storage an opt-in feature rather than the default. Customers can check their settings and choose the level of data retention they prefer.
Why it matters - Risks of prompt and credential leaks in AI development
This incident highlights the risks involved in handling customer prompts and credentials in AI development. Prompt leaks can expose sensitive personal or business information. Such leaks can damage customer trust and lead to legal or regulatory issues.
Credential leaks pose a different but related risk. If an attacker gains access to stored prompts or other data, they can search for secrets or credentials. These secrets could be reused to access other systems, increasing the scope of the breach.
The incident shows how quickly an exposed credential can be exploited. Even after reverting a commit, the credential may remain active for a short window. This makes it essential to use security measures like short-lived credentials and zero data retention policies.
For organizations developing AI tools, this case emphasizes the importance of security controls. Regular audits, prompt revocation, and careful handling of sensitive data are vital. Protecting customer data helps maintain trust and avoids costly breaches.
How OpenSmartRoute helps - Routing requests to improve security and control
OpenSmartRoute is an open-source router for AI requests. It helps teams improve security by controlling how requests are handled. Each request is sent to the best-fit model, agent, or tool from a catalog defined by the team.
The router scores each candidate based on quality, cost, speed, and safety. The team sets the importance of each factor for every request. Hard rules ensure that personal data stays on an on-premises model or region. The router learns from outcomes, giving more traffic to models that perform well.
Using OpenSmartRoute, teams can add new models easily. A new model is simply added as a catalog entry and starts competing on requests. The system also shows the cost of each request compared to the most expensive option. It includes an input guard to detect prompt injection and personal data before requests leave.
This setup helps teams reduce risks associated with data leaks. Routing requests through OpenSmartRoute allows better control over where sensitive data goes. It also provides transparency and accountability for model use and costs.
How it compares - what existed before, what this changes and what stays the same
Before this incident, many AI development teams relied on static credentials stored in code repositories or environment variables. These credentials often had long lifespans and broad access rights. Developers used version control systems to manage code, but credentials could be accidentally committed or exposed during development.
Many systems used manual processes to revoke or rotate credentials after a leak was suspected. This approach was slow and often incomplete. It depended on human intervention and could leave windows of vulnerability open.
This incident shows that reverting a commit does not necessarily stop credential misuse. Even after the commit was reverted, the credential remained active during the exposure window. Attackers can still use the credential until it is revoked or expires.
The change introduced by this event is the emphasis on short-lived, scoped credentials and zero data retention policies. Together AI now defaults to zero data retention for new accounts. This limits the amount of prompt data stored and reduces the impact if a leak occurs.
What stays the same is the need for ongoing security practices. Regular audits, credential checks, and careful handling of sensitive data remain essential. Automated detection and quick revocation are critical to minimizing damage.
The incident also highlights the importance of layered security. Using keyless access where supported, and implementing credential checks on pull requests, can prevent accidental leaks. These measures complement traditional security controls.
In summary, the industry is moving toward more dynamic, short-lived credentials and stricter data policies. These changes aim to reduce the risk of leaks and limit their scope when they happen. However, fundamental security practices remain vital.
Questions this leaves open - what the source does not say and how a reader can check it
The source does not specify how long the credentials remained active after the commit was reverted. It also does not detail the exact methods used by the attacker to test the validity of other stored credentials.
It is unclear whether the attacker used automated tools to search for credentials or manually tested them. The extent of the attacker’s activity beyond the initial download remains unknown. Did they attempt further access or data extraction?
The source does not mention how many other credentials or secrets might have been stored in prompts or code repositories. It also does not specify whether other customer data was exposed beyond the samples downloaded.
Readers can check their own security posture by reviewing how they manage credentials. Are they using short-lived, scoped tokens? Do they revoke credentials promptly after a leak? Are they monitoring for unusual activity?
Organizations should audit their stored secrets regularly. They can also implement automated tools to detect leaked credentials or sensitive data in code repositories and prompts. Using zero data retention policies can limit exposure.
It is advisable to verify whether their systems support keyless access or short-lived credentials. If not, they should consider adopting these practices. Regular security reviews and incident response plans are also recommended.
The incident underscores the importance of transparency. Companies should disclose how they handle credential leaks and what measures they take to prevent future incidents. Customers should ask about their security policies and data handling practices.
Finally, organizations can stay informed by following updates from security researchers and industry groups. They can participate in security audits and share best practices to improve their defenses against similar incidents.