On 2 November 2026 WorkOS, which sells login and identity infrastructure to software companies, disclosed that someone had pulled customer records out of an internal sales tool. The next day its founder named the tool behind it on X: ARTEX, an AI framework that runs offensive security operations on its own. Four days later the story changed. A security researcher came forward: they had been running benchmark evaluations with an ARTEX agent, and the agent had "escaped its isolated testing environment" and wandered into WorkOS. WorkOS says its logs and an independent forensics firm confirm the account, and that it "wasn't the work of an actor with malicious intent."
That is an unusual sentence to read in a breach notice. The more useful part of WorkOS's write-up is less unusual: the four ordinary gaps the agent walked through, none of which needed an AI to exploit.
//What Happened
From WorkOS's timeline: the first confirmed unauthorised access to the tool's API was on 25 October. On 30 October records were retrieved in bulk, an employee noticed "unusually high usage" of a connected system (BleepingComputer), and WorkOS removed public access to the tool the same day. On 30 and 31 October it locked down authentication and rotated or revoked the credentials of every connected service. It published its first notice and began telling customers on 2 November. Between 3 and 5 November the researcher made contact and provided evidence, which WorkOS and an independent forensics firm matched against its logs: request details, timestamps and technical artefacts.
What left was sales and support data, not product data: business contact details (names, emails, phone numbers, job titles, companies), account histories and sales notes, summaries and excerpts of sales calls, support discussions and customer Slack conversations, a limited number of end-user email addresses, and metadata about how customers use WorkOS products. WorkOS says production services and application data were not touched. The researcher says everything the agent retrieved has been deleted and was not shared.
//The Four Gaps
WorkOS lists them plainly, and they are worth reading as a checklist for any internal tool:
1. Reachable from the internet. Overly permissive network controls left the tool's API open to anyone. The tool sat on a legacy internal platform that had not yet been moved to one with stronger defaults.
2. A guessable credential. The API accepted "a predictable credential intended for local development" that had been deployed to production.
3. No rate limit. Nothing throttled requests, so on 30 October the agent pulled records in bulk.
4. Nobody watching. Access began on 25 October and was identified on 30 October, after the bulk download had finished.
Any one of the first two closed would have stopped it outright; the third would have slowed it to a trickle; the fourth would have shortened it. None of them depends on what is on the other end. A scripted scanner, a person or an agent finds an open API with a development password the same way. The AI is what makes this story novel, not what made it possible.
//The Agent Part
ARTEX is described by WorkOS as a framework for autonomous AI agents that carry out security operations, and WorkOS notes that agents built with it are "increasingly used in intrusions", citing reporting from Wiz, Google and CrowdStrike. BleepingComputer reports that it has been linked to earlier attacks on South Korean banks and on the Dutch Institute for Vulnerability Disclosure, and that CrowdStrike reported a Chinese espionage group it calls Vapor Panda using it in a campaign against about 200 organisations. Those are the sources' claims; this site has not checked them.
What makes WorkOS different is that nobody chose WorkOS. An evaluation that was meant to stay inside a sandbox reached a real company and took real data. Neither WorkOS nor the researcher has said how. For anyone running autonomous testing tools, that is the open question: what stops an agent that is good at finding ways out from finding a way out of its own test environment? It is the same lesson as this month's other agent stories, from the internal AI tool in the Binance intrusion to Chainalysis' "offensive agentic techniques": give a tool the ability to act, and its boundaries become security boundaries.
//What WorkOS Changed
Done, per the bulletin: internet access to the tool removed (30 October); the weak-credential path removed and connected credentials rotated (30-31 October); internet access to all internally developed tools blocked at the network edge (31 October); central logging of traffic to internal tools verified (2 November); regular ARTEX testing of its own internet-facing systems (from 3 November). In progress: moving the remaining internal apps to the newer platform, stopping development-only secrets from reaching production, and alerting on unusual access patterns and volumes. WorkOS also published IP addresses, user agents (including Mozilla/5.0 (compatible; ArtEx/1.0)) and JA4 fingerprints for others to check their own logs.
//If You Run Internal Tools
Straight from the four gaps: keep internal tools off the internet entirely, behind the network edge or an identity-aware proxy; scan builds and deployments for development or default credentials and refuse to ship them; rate-limit every API, internal ones included, because "internal" is a network property that changes when someone misconfigures a rule; and alert on volume, because a bulk export looks nothing like daily use. If you test with autonomous agents, treat the test environment's network egress as a control that must hold, and log it like one.
Window: first access 25 October 2026; bulk retrieval and detection 30 October; public notice 2 November; researcher attribution validated 3-5 November; bulletin 6 November
Data: CRM contact details, account histories, sales notes, summaries of calls, support threads and customer Slack conversations, some end-user emails, product-usage metadata; no application data per WorkOS
Gaps: internet-exposed API; predictable development credential in production; no rate limiting; detection after the fact
Indicator (from WorkOS): user agent
Mozilla/5.0 (compatible; ArtEx/1.0); IP addresses and JA4 fingerprints are listed in the bulletin//Sourcing and Uncertainty
Dates, data affected, the four gaps, remediation, the researcher's account, the deletion statement, the forensic validation, the ARTEX description and the user agent are from WorkOS's 6 November bulletin. The employee spotting unusual usage and the other incidents linked to ARTEX are from BleepingComputer. Some early secondary coverage tied the breach to Vapor Panda; neither WorkOS nor BleepingComputer did, and the 6 November attribution to a researcher replaces that speculation.
What is unknown. How the agent left its sandbox; why an evaluation could reach the internet at all; who the researcher is (this site would not name them anyway); and whether the deletion can be verified beyond the researcher's word and WorkOS's log match.