AI Security
WorkOS was breached through an internal sales tool. Ask these four questions about yours
WorkOS says an AI agent that escaped a researcher's sandbox reached an internal go-to-market tool that was exposed to the internet, accepted a development credential in production, had no rate limits, and went unnoticed for five days. Each root cause becomes a question HealthTech vendors can ask about their own internal tools this week.
November 11, 2026 · 6 min read

An internal tool that was meant to stay private was reachable from the internet, accepted a predictable development credential in production, and had no rate limiting. That is how an AI agent pulled customer records in bulk from WorkOS, according to the company's Nov. 6 security bulletin. WorkOS says the agent belonged to a security researcher running benchmark evaluations with ARTEX, and that it escaped its isolated testing environment and reached WorkOS unintentionally.
The attribution is the headline. The root cause section is the useful part. WorkOS published four specific gaps, and none of them is about AI. They are about an internal application that never got the same controls as the product.
What WorkOS says happened
The affected system was an internal application WorkOS's go-to-market teams use to look up and manage account information. Per the bulletin, it aggregates data from the CRM and from customer support and communication systems. It sat on a legacy internal tools platform. Most internal tools had already moved to a newer platform with stronger baseline controls. This one had not.
The timeline in the bulletin:
- Oct. 25: first confirmed unauthorized access to the tool's API.
- Oct. 30: records retrieved in bulk. WorkOS removed public access and began remediation.
- Oct. 30 to 31: authentication secured, credentials for connected services revoked or rotated, access to all internally developed tools restricted.
- Nov. 2: initial public post and customer notification began.
- Nov. 3 to 5: the researcher came forward with evidence, an independent forensics firm validated the attribution, and the researcher confirmed the data was deleted.
The exposed data included business contact details from the CRM, account and relationship history, operational metadata about product usage, and summaries and excerpts from sales calls, support discussions and customer Slack conversations. WorkOS says a limited number of support communications included end-user email addresses.
BleepingComputer notes that neither WorkOS nor the researcher has publicly explained how the agent escaped its testing environment. I am not going to speculate on that. The containment failure on the researcher's side is a separate lesson, and I wrote about a similar one in OpenAI's sandbox incident. This post is about the receiving end.
Why this maps to HealthTech vendors
Look at the data list again and replace "WorkOS" with any vendor selling into hospitals. The CRM holds health system contacts, titles and phone numbers. The support system holds ticket threads, sometimes with screenshots. Call summaries describe a customer's environment and problems. Shared Slack channels with customers carry whatever people typed in them.
Security questionnaires from health systems ask hard questions about the product. They rarely ask about the admin console the sales team built to see everything about an account in one place. I made the same point about analytics tools that never made the inventory. Internal tools that join several data sources into one view are some of the most valuable systems you run, and often the least reviewed.
The four-question internal tools check
Each question comes straight from one of the four gaps in the WorkOS bulletin. Run it against every internally built tool that reads customer data, starting with the ones that aggregate more than one source.
-
Can it be reached from the internet? WorkOS says the tool "should not have been reachable from the internet" and that overly permissive access controls exposed its API endpoints publicly. Ask the question for the API, not only the web interface. A login page behind single sign-on means little if the API behind it answers anyone. The check is simple: list the tool's hostnames and endpoints and confirm from outside your network that they do not respond. WorkOS's own fix was to block internet access to all internally developed tools at the network edge.
-
Could a development credential work in production? The bulletin says the API accepted a weak credential intended for local development, and that controls did not stop it from being deployed to production. Ask how your build and deploy pipeline would catch that. Secret scanning finds credentials in code. It does not always catch a development-mode authentication path left enabled in a production configuration. WorkOS says it is adding controls to stop weak or development-only secrets from reaching production. Write down which control would do that job for you, and if the answer is "code review," treat that as a gap.
-
What stops bulk extraction after login? WorkOS says the API had no rate limiting or throttling, so once authenticated the agent extracted records quickly. Rate limits are an easy control to skip on internal tools because they have few users. That is the reason to add them. A sales tool used by a few dozen people has a predictable request profile. Set limits per user and per credential well above normal use and far below a full export.
-
Would you notice in hours, or in days? The bulletin's fourth gap is telemetry. First access was Oct. 25. WorkOS identified the activity on Oct. 30, after the bulk retrieval had ended, when an employee noticed unusually high usage of a connected system. Five days is a long time for an unknown client to be talking to a tool full of customer conversations. Ask whether logs from this tool reach your central logging, whether anyone alerts on request volume or new client sources, and whether the alert would route to a person who would act. WorkOS lists verifying centralized logging of network traffic to internal tools and adding alerts for unusual access patterns and request volumes among its fixes.
What to do with the answers
Put the results in a short table: tool name, data sources it joins, and a yes or no for each of the four questions. Any tool with a "yes, it is reachable" or "no, we would not notice" moves to the top of the list.
Then decide which platform these tools should live on. WorkOS's underlying problem, by its own account, was one tool left on a legacy platform after the rest had moved to one with baseline controls. If you have a standard platform for internal tools, the fastest fix is often migration rather than patching each tool. If you don't have one, the four questions are a reasonable starting definition of what it must provide.
Finally, update your questionnaire answers. If a health system asks where its contacts and support history live, the honest answer now includes these tools.
WorkOS says it has started testing its own internet-facing systems regularly with ARTEX, looking for exposed systems and guessable credentials. Which of your internal tools would turn up first if someone ran that test against you tomorrow?