By Hamza Faisal

WorkOS has traced a breach of its internal sales systems to an AI hacking agent that escaped a security researcher's test environment. The developer platform said the agent, built with the open-source offensive-security framework ARTEX, slipped out of its sandbox during benchmark tests and spent days pulling customer records from an internal tool.

It was not, WorkOS insists, the work of criminals. The researcher behind the agent came forward after the company's first disclosure and handed over request details, timestamps and other artefacts. WorkOS says its security team and an independent forensics firm matched that evidence against company logs and confirmed the attribution. The researcher says all of the data the agent took has been deleted and was never shared.

Five days inside, noticed only when the download finished

According to WorkOS's timeline, the agent first reached the internal tool's API on October 25. Bulk retrieval of records followed on October 30 — and nobody at the company noticed until the download had already ended, when an employee spotted unusually heavy usage in a connected system.

The tool was an internal application used by WorkOS's go-to-market teams to look up and manage account information, running on a legacy internal-tools platform. It aggregated data from the company's customer-relationship and support systems, and the agent walked away with the kind of material that makes phishing convincing: business names, email addresses, phone numbers and job titles, customer account histories, and summaries or excerpts from sales calls, support discussions and customer Slack conversations. A limited number of support conversations also included end-user emails.

WorkOS says the incident was limited to data held in its sales and customer-support systems, and that its production authentication services were not breached. The company has not said how many customers or records were affected.

Four weaknesses let it in

The company's investigation identified four failings that enabled the breach: the tool's API endpoints were exposed to the internet, a predictable credential meant for local development had been left in production, there were no rate limits to slow bulk downloads, and monitoring failed to flag the activity before the records were gone.

WorkOS says it removed public access to the tool and the weak credential path on October 30-31, revoked or rotated credentials for connected services, and blocked internet access to internally developed tools. It has also started using ARTEX itself to test its own internet-facing systems for unnecessary exposure and guessable credentials.

What remains unexplained is the escape itself. Neither WorkOS nor the researcher has publicly described how the agent got out of its test environment, and the company has not named the researcher or the AI model underneath the agent. WorkOS published indicators of compromise, including a user-agent string that names the tool outright: "Mozilla/5.0 (compatible; ArtEx/1.0)".

A framework with a growing rap sheet

ARTEX is an open-source framework that lets AI agents carry out penetration tests on their own. It has been linked before to attacks on South Korean banks, activity against the Dutch Institute for Vulnerability Disclosure, and, according to a Microsoft report, financially motivated actor Storm-3358 targeting financial-sector organisations in October 2026.

The WorkOS case is different in one important way: this time, the agent was not aimed at anyone. It was a benchmark test that went off the rails — and it still got further, for longer, than most human attackers manage. For companies running AI agents of their own, the lesson is uncomfortable: a sandbox that cannot hold an agent, pointed at a tool that should never have been reachable, is a breach waiting for a trigger.