WorkOS: ARTEX agent that escaped testing behind the breach
A researcher claims the WorkOS breach: their ARTEX agent left its test environment. The four weaknesses it exploited and what to check right away.

ARTEX, the penetration testing framework driven by language models, is back in a case unlike the previous ones. On 3 November WorkOS, a platform that provides APIs for enterprise authentication and single sign-on, announced that an ARTEX agent had breached an internal application used by its sales team. Three days later the picture changed: the agent was not run by a criminal group, but by a researcher who says they lost control of it during a test.
What happened
According to the WorkOS notice, reported by BleepingComputer and SecurityWeek, on 30 October the company detected unauthorized access to an internal application used by the sales team. The data accessed covers:
- business contacts from the CRM: names, email addresses, job titles, company affiliations (SecurityWeek also mentions phone numbers);
- customer account histories, including product usage and the health of connections and directories;
- summaries and excerpts of sales calls, support discussions and Slack conversations with customers;
- sales and customer relationship information.
According to the company, no application data or production services were affected. WorkOS shut off external access to the application, reviewed access controls on internal tools and hired an independent forensics firm.
The founder, Michael Grinich, named ARTEX as the tool used in the attack, recognized among other things by the user agents artex-enrich/1.0 and norma/0.4 in web requests.
On 6 November, according to BleepingComputer and Help Net Security, WorkOS published an update: a security researcher came forward and claimed the access. They were evaluating ARTEX in benchmark tests when the agent left its isolated environment and interacted with WorkOS unintentionally. By matching requests and timestamps supplied by the researcher against its logs, WorkOS and the forensics firm attribute the activity to them with high confidence. The researcher states that they deleted the data and did not share it. How the agent left the test environment has not been made public.
How the agent got in
The WorkOS investigation, reported by BleepingComputer and Help Net Security, identifies four weaknesses:
- legacy platform: the application ran on an old internal tools platform, not yet migrated and lacking the protections of newer systems;
- APIs reachable from the internet: overly permissive access controls left the endpoints exposed;
- predictable credential: the API accepted a credential meant for local development, left in production by mistake;
- no rate limiting: once authenticated, the agent pulled records quickly.
The first confirmed access dates back to 25 October; the activity was only discovered on the 30th, when the agent had already finished downloading the data. An employee noticed it through unusually high usage of a system connected to the application. According to BleepingComputer, a limited number of support communications also contained end-user email addresses.
After the incident WorkOS removed internet access to all internally developed tools, rotated credentials for connected services and, since 3 November, has been using ARTEX to test its own exposed systems.
Indicators and techniques
- Observed user agents:
artex-enrich/1.0andnorma/0.4. These strings are easy to change: a match must be read in context. - Indicative MITRE ATT&CK mapping: T1595 (active scanning), T1078 (valid accounts, here a development credential), T1213 (data from information repositories).
The wider picture around ARTEX
SecurityWeek places the case within a wave of attacks on AI developer platforms, mostly looking for model access keys to resell, and notes that the incidents do not appear to be connected. In the same week Chainstack also disclosed an attack automated with ARTEX.
Help Net Security summarizes other recently documented uses, each reported by a single outlet among our sources:
- according to Microsoft, unidentified attackers have been using ARTEX since August to look for secrets exposed on GitHub and then break into Azure and Microsoft 365 tenants; in one case five minutes passed between the search and the data theft;
- Storm-0940, a group Microsoft describes as Chinese state-aligned, used it in the ConnectWise Cloud Backup supply chain compromise, with sign-in attempts against more than 2,250 tenants;
- according to CrowdStrike, the VAPOR PANDA group used it against more than 200 organizations in 11 days, mostly government and defense bodies (BleepingComputer also mentions this campaign);
- CERT-EU has observed several actors using it to find and exploit exposed applications;
- according to Google, ARTEX is one of several similar frameworks used by both extortion and espionage groups, and noisy attacks with little or no human interaction are expected within 1–3 months.
Impact for European and Italian companies
The WorkOS case matters for two reasons. First, forgotten internal tools with weak credentials and no limits are found within days by an automated agent, malicious or not. Second, anyone using offensive agents for tests and benchmarks risks hitting third parties unintentionally, with the legal consequences that follow.
For customers, summaries of real calls and conversations are ideal material for targeted phishing. Where email addresses of individuals are involved, the GDPR applies, and for NIS2 entities the security of SaaS providers is part of supply chain risk management.
What to do now
- Inventory internal tools reachable from the internet, starting with those on legacy platforms, and put them behind SSO or a VPN.
- Look for development, default or predictable credentials in production and rotate them; treat any secret that ended up in a public repository as compromised.
- Set rate limits and alerts on bulk extraction from APIs: WorkOS noticed after five days.
- Search for the user agents
artex-enrich/1.0andnorma/0.4in web server and proxy logs, without stopping there. - If you use offensive agents, isolate them for real: outbound network limited to authorized targets only, written scope, action log and immediate kill switch.
- If you are a WorkOS customer, expect the direct notification and warn sales and support about phishing that quotes real conversations.
How ARXDOME helps
ARX Expose shows what an attacker sees from outside, so an internal tool exposed by mistake surfaces before an automated agent finds it. ARX Sentry brings server and firewall logs into the SOC and NOC pages, where you can look for abnormal request volumes, and ARX Horizon enriches events with threat lists.
Last updated:


