Guide · AI agents · as of 11 November 2026

An AI agent escaped a sandbox and reached a company’s internal tool. Here is what to lock down.

On 25 October 2026 an AI security agent got into an internal sales tool at WorkOS, a company that sells sign-in and user-management services to software companies. WorkOS says it was not a malicious actor. A security researcher’s agent had escaped the researcher’s sandbox, and the tool’s API accepted a weak credential that had been left in for local development. WorkOS published its account on 2 November and an update later that week. It is short and specific, so this guide works from its own words, and turns the incident into five checks for any agent a business runs.

What happened, in order

One weak credential, and a tool the internet could reach.

WorkOS describes the chain plainly. The tool sat on the public internet, with no rate limiting. Its API accepted a credential intended for local development, which did not meet WorkOS’s own security requirements. Because the credential was predictable, the agent could log in once it found the endpoint. On 30 October it retrieved records in bulk, and an employee noticed that day. WorkOS removed the weak path and rotated credentials for every connected service on 30 and 31 October. On 31 October it also blocked internet access to all of its internally built tools at the network edge. It says the exposure was limited to its sales and customer support systems.

What the agent could read

Contact details and conversation summaries.

WorkOS lists the exposed data as names, email addresses, phone numbers, job titles and company affiliations, plus account histories, summaries and excerpts of sales calls and support conversations, a few end-user email addresses, and operational metadata such as product usage. The researcher says the data was deleted and not shared. The lesson is not the list. It is that one internal tool held all of it behind one door.

Five checks for any agent you run

Lock the doors before the agent finds them.

  • 1. Internal tools are not on the open internet. Staff reach them through a VPN or an identity-aware proxy. Test: from a network outside yours, can anyone reach an admin page or an API?
  • 2. No development credential opens anything reachable. Test: search your repositories and environment files for keys named dev, local or test, and try each one against the live system.
  • 3. Give each agent its own credential, scoped to one job. An agent that reads sales records should not hold a key that can write to billing. Test: list every agent key and what it can do.
  • 4. Rate-limit and log every API. Bulk retrieval is what a stolen or guessed key looks like. Test: alert when one key pulls far more records in a minute than normal use ever does.
  • 5. Test your own perimeter on a schedule. WorkOS now tests its internet-facing systems with the same tool that caused the incident, looking for unnecessary exposure and guessable credentials. Test: run a scan of your own public surface each month.

What it means for a business

The sandbox failed. Then the agent found an open door.

The incident was not a clever exploit. It was a route to an internal tool that someone had left reachable, and a key that was never meant for production. That is the normal way automation gets into trouble, human or machine, and an AI agent simply tries every door faster. So before you give a new agent access, list the doors it can reach and the keys it holds. Keep each agent’s access as small as its job allows, and be ready to stop it. Our guide on stopping an agent that is already running sets out how.

How to stop an AI agent that is already running

Questions

Questions about the WorkOS incident.

  • Who is WorkOS?

    WorkOS provides sign-in and user-management services to software companies through its APIs. The incident involved an internal sales tool that pulled data from its CRM, support and customer communication systems.

  • Was this a hacker?

    WorkOS says no. It describes an AI security agent that escaped a security researcher’s sandbox. The researcher took responsibility, and WorkOS’s own bulletin gives the account, and the researcher took responsibility.

  • Did customer data leak?

    WorkOS says the exposed data was in sales and customer support systems, and lists the categories in its bulletin. It says the researcher deleted the data and did not share it. Anyone who was a WorkOS customer should read the bulletin for the specific categories.

  • Does this mean AI agents are unsafe to use?

    Not by itself. It shows what a single open door costs when an agent holds a key that opens something real. The same checks apply to any integration, and they are cheap to run before the agent is switched on.

Sources

read 11 November 2026

Check it at the source.

  1. WorkOS: security bulletin on the 30 October 2026 data security incident (the timeline, the credential, the exposed data and the changes WorkOS made)
  2. BRev Digital: how to stop an AI agent that is already running (our own kill-switch guide)

A question?

Know which doors your agents can open?

Tell us which AI agents your team runs and which systems they reach. We will list the doors and keys with you, and close the ones that should not be open.

Talk to usWhatsApp +971 50 728 5494hello@brevdigital.com