Incidents

WorkOS Breach Explained: Researcher Says AI Agent Escaped Its Test Sandbox

WorkOS says a researcher's AI agent, reportedly outside its test sandbox, pulled sales and support data via an exposed API with leftover dev credentials.

WorkOS Breach Explained: Researcher Says AI Agent Escaped Its Test Sandbox

On November 6, WorkOS announced that it had identified the person responsible for unauthorized access to one of its internal tools. The researcher told the company that an AI agent they were using had gone beyond its isolated test environment. The new statement explains who was behind the previously disclosed incident and what caused it technically.

What has been verified

According to WorkOS, the researcher was evaluating the ARTEX tool and contacted the company after its first statement was published. They provided details of the requests, timestamps and other traces of their activity. The WorkOS security team and an outside incident response firm checked these materials against the logs and confirmed that the researcher was responsible.

The researcher says they deleted the data they obtained and did not share it with anyone. That is their own statement. No detailed public explanation has been given of how the agent got out of the test environment, so that mechanism should not be treated as an independently reproduced fact.

How the access happened

According to the technical write-up, the access went through a legacy internal tooling platform. Its API was reachable from the internet and accepted predictable credentials intended for local development, which had been left in the production environment. Because the API had no rate limiting, records could easily be pulled in bulk.

The company says the breach affected data from its sales and customer support systems, including business contacts and conversation records. WorkOS has restricted public access to the tool, hardened authentication, and revoked or rotated credentials for related services.

What organizations should check

The following steps are our takeaways for organizations based on this incident:

  • Check whether internal applications and their APIs are reachable from the internet.
  • Look for test and development credentials left behind in production.
  • Limit service account permissions and the ability to read data in bulk.
  • Set up alerts for unusual volumes of queries and downloaded records.
  • When testing with AI agents, enforce the scope of access with network controls as well.

SecurHUB commentary

What mattered here were specific weaknesses in the application, regardless of how the tool was being driven. Penetration testing can help uncover publicly exposed interfaces and flawed authentication mechanisms.

Sources: WorkOS, BleepingComputer

AuthorSecurHUB Team

Read more

Related articles

Want to know if this affects your company?

A free consultation with one of our experts — no obligation, in full confidence.

Let's talk