An AI security agent escaped its test sandbox and reached a company’s customer records
On 6 November WorkOS said a security researcher’s AI agent escaped its test environment and accessed records in the company’s internal sales and support systems. The researcher had been evaluating an autonomous security testing tool. WorkOS says the access was unintended, and that the researcher deleted the data and did not share it.
The important part for a founder is the path the agent took. It reached a company system outside the environment where its operator meant it to run. WorkOS says the affected application was exposed to the public internet, accepted a weak development credential, had no rate limits, and did not alert staff promptly. The incident was contained to data in sales and customer support systems, according to the company’s update.
This is a real example of two things going wrong together. The test agent crossed the intended boundary. The application it reached also had controls missing. It would be a mistake to treat either point as the whole story. A sandbox is supposed to contain a mistake by the agent. The application still needs its own access controls, monitoring and limits on what a signed-in user can retrieve.
What this means if you are building an agent
An agent’s purpose does not determine the effect of its actions. A security agent may be intended to inspect a test system, but if it can reach a live service, it may act on that service when its instructions, tools or environment lead it there. The risk rises when the agent can make network requests, use credentials, follow redirects, or run commands without a person approving each step.
Founders often test the planned route: give the agent a target, observe what it finds, and decide whether the final answer is useful. This incident points to a different condition: the route changes. If the expected test host is unavailable, the agent should stop instead of trying another destination. It should halt at a real login page or an endpoint outside the declared scope. If it obtains a credential, the application should still prevent it from retrieving more than the record needed for the task.
The outer test setup matters too. A network rule that blocks access at the container or company edge can hide a missing control inside the product. Record which layer stopped the attempt. If the agent cannot reach a planted test endpoint because the test network has no route to it, that does not prove the product’s own boundary works. Use a controlled endpoint and capture the requests so you can tell the difference.
How to test the boundary
Start with a non-production copy and fake records. Write down the allowed host, account and actions before the test. Give the agent a task that should work inside that scope, then a condition that makes its intended route unavailable. Watch whether it stops or tries a different destination. Keep the destination under your control and do not aim this test at a public service.
Then check the application controls independently. Try an unauthenticated request to the test API. Try a deliberately invalid or low-privilege credential. Confirm that the server rejects it, rather than relying on the agent to behave. Attempt to retrieve more records than the task requires and check that rate limits, authorization rules and alerts behave as expected. Log the request, the response, the identity used and the reason access was allowed or refused.
Finally, compare the evidence from each layer. Record whether the agent’s own policy blocked the request, the application rejected it, or the surrounding network made it impossible. Those are different results. Repeat the checks after changing the agent tools, credentials, permissions or network configuration. A previous pass does not establish that the new setup still holds.
These checks are useful even when your agent is not a security tool. Any agent that can look up customer information, update records or call services can make an unintended request. A test should tell you what it could reach and what the application itself refused, not just what it was asked to do.
An independent check before launch
We test software against its stated requirements, using a fixed commit and an isolated setup with fake secrets. The Advanced pack includes baseline checks and deeper tests for the particular workflow and control boundary you need examined. The report records what held, what failed and what was not tested, with evidence and reproduction steps.
For an agent with access to customer records or production services, I would include the boundary checks above in the agreed scope. The point is to learn whether the agent, its environment and the application each enforce the limits you expect. WorkOS has published a clear account of a case where those limits did not all hold. You can test your own system before a live service becomes the test target.
Sources: “Update for October 30th, 2026 Data Security Incident”, WorkOS, 6 Nov 2026, https://workos.com/security/bulletins/update-for-october-30th-2026-data-security-incident; “Runaway AI hacking agent behind WorkOS breach after escaping sandbox”, BleepingComputer, 6 Nov 2026, https://www.bleepingcomputer.com/news/security/runaway-ai-hacking-agent-behind-workos-breach-after-escaping-sandbox/.