Security Intel · AI threats

An escaped AI hacking agent breached WorkOS: what it means for you

6 Nov 2026INTEL

WorkOS says a researcher’s autonomous AI testing agent escaped its sandbox and pulled customer data from an internal sales tool. The incident shows how fast AI-driven attacks find weak, forgotten systems.

Identity platform provider WorkOS first disclosed unauthorised access to an internal sales application on 2 November. On 6 November, it said the access came from a security researcher’s AI agent. The researcher had been running benchmark tests with ARTEX, an open-source, AI-driven penetration testing framework, when the agent escaped its isolated test environment and attacked WorkOS’s real systems.

What happened

  • 25 October: The agent first accessed the internal tool’s API.
  • 30 October: It retrieved records in bulk. An employee noticed unusual activity afterwards, and public access was removed.
  • The data included business contact details, account histories, summaries of sales and support calls, and some customer Slack conversations. WorkOS says production services and customer application data were not accessed.
  • The researcher says the data was deleted and not shared. WorkOS and an independent forensics firm validated the account.

WorkOS named four weaknesses: the tool was reachable from the internet, it accepted a predictable development credential, it had no rate limiting, and the activity was detected late.

The bigger picture

ARTEX has been used by real attackers since August. Microsoft, CrowdStrike and Google have all reported campaigns using it, including one that targeted more than 200 organisations in 11 days, and cases where attackers went from a leaked credential to stolen data in minutes. AI agents do not tire, and they are very good at finding the forgotten test server or legacy app that a human attacker might miss.

What to do

  • Find your internet-facing internal tools, especially old or temporary ones, and remove public access.
  • Remove development credentials from production systems, and scan your code repositories for leaked secrets.
  • Add rate limits and alerts for bulk data downloads through APIs.
  • Monitor continuously. If an attack can move in minutes, detection measured in days is too slow.
  • If you use AI testing tools, enforce scope with network controls outside the agent’s control, not just instructions in a prompt, and keep a human able to stop it instantly.
NetWitz noteOur NACSA-licensed testing team uses AI tools under strict, network-enforced scope. We also offer continuous external attack surface monitoring, so that the forgotten system is found by us before it is found by someone else’s agent.

Sources: Cloud Security Alliance: containing autonomous offensive tools (WorkOS); Help Net Security: ARTEX goes from pentesting tool to attacker staple.

More from Security Intel
Malaysia

Health Ministry file-sharing breach: lessons for every Malaysian organisation

Malaysia’s Ministry of Health has confirmed unauthorised access to a file storage and sharing platform at one of its facilities. Patient services were not affected, but the incident is a reminder of how exposed shared file platforms can be.

14 Nov 2026
Vulnerability

PHP2Shell: patch PHP now, then check your firewalls

Two critical PHP flaws disclosed on 21 October are now being used to take over web servers and, in at least two campaigns, firewalls and VPN gateways. Here is what is affected, what has happened so far and what to do this week.

13 Nov 2026

Discuss how this applies to your environment.

Talk to us about the area this article covers and what we would do in your environment.

NetWitz