Collectively AI disclosed on November 21, 2026 that an unauthorized get together downloaded saved samples from a subset of buyer prompts submitted to its serverless inference service. An AI coding agent had uncovered the private-repository credential used within the obtain whereas the corporate was creating its public xorl GitHub repository.
Prospects had approved the prompts for routine software program testing, and Collectively AI saved the samples in a personal Hugging Face repository. They concerned requests to the MiniMax M3, GLM-5.2 and Kimi K2.6 fashions, totally on June 24, 2026, July 4, 2026, July 8, 2026 and July 14, 2026.
Credential Publicity and Unauthorized Exercise
The coding agent dedicated a credential that granted entry to the non-public Hugging Face repository throughout xorl’s growth. Collectively AI’s automated scanning detected the publicity, and the corporate reverted the commit. Throughout that publicity window, an unauthorized get together obtained the credential and used it to obtain the saved samples.
The sequence makes clear that detecting and reverting the commit didn’t forestall the credential from getting used in the course of the publicity window. The mannequin names set up which requests had been sampled, whereas the repository credential explains how the saved prompts grew to become accessible.
After the obtain, the get together searched the samples for credentials to different providers and examined whether or not they had been legitimate. 5 of the examined credentials had been canary credentials related to a Dragos safety experiment. The searches and validity checks present that the exercise prolonged to attainable reuse of secrets and techniques contained in prompts, making a danger past disclosure of the immediate textual content itself.
Investigation and Buyer Affect
Collectively AI labored with an exterior forensic investigation agency and regulation enforcement. The corporate stated its evaluation has not recognized proof of further immediate leakage past this incident, that it has no motive to imagine the information was leaked additional, and that it understands the information has now been deleted.
Buyer affect nonetheless will depend on the content material of the samples. The evaluation is ongoing to find out whether or not delicate private data or different buyer information was uncovered, and Collectively AI will contact affected clients instantly as wanted with findings related to them.
Credential and Retention Modifications
Collectively AI’s response consists of scoped or short-lived credentials and keyless entry the place supported, credential checks on all pull requests, and up to date growth processes governing using AI coding brokers. The corporate stated the credential measures are designed to cut back the potential affect of publicity, whereas the opposite modifications add controls round pull requests and coding-agent growth.
Zero information retention, abbreviated ZDR, is now the default setting for brand spanking new accounts. Collectively AI recommends that clients dealing with delicate immediate information allow ZDR of their account settings and offers safety@collectively.ai for questions concerning the incident or its potential affect on a company.
The incident connects credential security in AI-assisted growth with the retention of buyer prompts for testing. Collectively AI has documented controls for each, however the ongoing content material evaluation stays the decisive step in establishing what the publicity means for patrons and which notifications are obligatory.

