Safety & governance Industry development
Together AI discloses a prompt-sample leak and ongoing customer review
Together AI says an exposed repository credential let an unauthorized party download stored samples of inference requests from June and July 2026. [1]
- Development status
- Company disclosure published November 21; review of affected prompt contents ongoing [1]
What happened
In a November 21 disclosure, the company said an AI coding agent committed a credential to a public code repository. Together says it removed the commit after detection, but the credential had been used to access a private repository containing prompt samples. Its review of affected content and customer notifications was continuing. [1]
Why it matters
The disclosure connects AI coding workflows, stored inference data and credential controls: removing a leaked credential from the current code does not establish that it was never used. [1]
Who is affected
Organizations that submitted potentially sensitive prompts to Together AI, and teams that allow coding agents to modify repositories or access development credentials. [1]
Key dates
- Together AI discloses access to stored prompt samples [1]
What it means in practice
Check account data-retention settings and relevant customer notices. Together recommends enabling zero data retention for sensitive prompt data and says it is making that setting the default for new accounts. [1]
Risks and open questions
This is the company's account of an investigation, not an independent forensic conclusion. The disclosure concerns stored samples from June and July, not a newly established November leak of all service traffic. The company had not finished determining whether sensitive personal information was exposed. [1]
Market relevance
Companies and organizations
Documented involvement in this development.
- Together AI
- Inference provider disclosing the prompt-sample incident [1]
Sector context
-
AI inference data governance and credential security Editorial assessment
Prompt retention and development credentials create risks that need separate review from model quality. [1]
Company-reported developments
-
Together AI's November 21 disclosure says an exposed repository credential enabled unauthorized downloading of stored samples from June and July inference requests. It describes control changes while stating that review of affected data remains ongoing. [1]
What to watch next
-
Editorial watchpoint
Check retention settings and subsequent incident findings relevant to the organization rather than assuming the disclosure proves complete remediation. [1]
Sources and verification
Catalog evidence checked How we verify records
Related profiles
Safety & governance
CCW autonomous-weapons instrument drafting mandate
States participating in the CCW Seventh Review Conference
Switzerland’s foreign ministry reports a drafting mandate following negotiations that concluded early on November21.
Safety & governance
Zscaler frontier-model security collaborations
Zscaler
The issuer says Claude Mythos 5.1 is integrated into Endpoint AI Security for its Exploit Paths capability, with a limited preview planned. It also describes a Mistral integration through its European sovereign Zero Trust Exchange.
Safety & governance
EU–UK Digital Partnership plan
European Union and United Kingdom
Paragraph54 of the joint statement identifies AI, including safety and security, among possible cooperation areas alongside connectivity, quantum technology, digital credentials and online safety.