Tensorlake, a company that builds sandboxes for running code written by AI agents, has disclosed a security incident in which an attacker reached its engineering GitHub repositories through vendor-hosted sandboxes. Founder Diptanu Choudhury disclosed the incident on 8 November 2026, saying the attacker pushed malicious code to branches and triggered GitHub Actions workflows to probe for secrets.

The attack began with an internal application that ran on an upstream provider and created sandboxes inside accounts used by Tensorlake's engineering team. Those sandboxes had access to the company's GitHub repositories and the automation workflows that build and deploy its code. Once inside, the attacker used that access to push code and to look for secrets available to the workflows.

According to the disclosure, the upstream provider had suffered its own security breach a couple of weeks earlier, and that breach exposed Tensorlake environment variables. Tensorlake had not rotated those variables because the application was not considered critical to production. Choudhury described that as an oversight, saying the company should have acted immediately.

Tensorlake says there was no impact on customer information such as API keys or sandboxes, and that it has no evidence any critical systems were accessed beyond the probing for secrets. No confirmed secret theft has been reported, and the disclosure gave no count of affected repositories or workflows. The company has rotated all of its secrets, removed the internal application from the upstream provider, and notified paying customers by email. Out of caution, it has asked all customers to rotate their Tensorlake API keys.

The incident is separate from the malicious Tensorlake npm package reported in October, when a compromised release shipped a credential-stealing worm. Further details are expected through customer communications and the company's trust centre.