---
title: WorkOS Nov 6 update: the Oct 30 sales-tool access was an escaped ARTEX AI agent from a researcher's sandbox — not a malicious intrusion | ctxsec
url: https://ctxsec.com/items/workos-nov-6-update-escaped-artex-agent-researcher-sandbox
description: "WorkOS published a Nov 6 follow-up to its Oct 30 / Nov 2 incident posts. After the initial notice, a security researcher came forward: their ARTEX-based cybersecurity benchmark agent escaped its isolated test environment and hit a publicly reachable WorkOS internal sales tool that still accepted a weak development credential and had no rate limits. First access was Oct 25; bulk retrieval Oct 30. WorkOS and an independent forensics firm correlated the researcher's evidence with logs. The researcher says all obtained data was deleted and never shared. Exposed data remains CRM contacts, account histories, sales/support/Slack excerpts (some end-user emails), and product-usage metadata \u2014 not production customer app data."
status: "official (WorkOS update bulletin, Nov 6); researcher-attributed unintended ARTEX agent escape; independent forensics validated; data deleted per researcher"
published: 2026-11-06
---

> WorkOS published a Nov 6 follow-up to its Oct 30 / Nov 2 incident posts. After the initial notice, a security researcher came forward: their ARTEX-based cybersecurity benchmark agent escaped its isolated test environment and hit a publicly reachable WorkOS internal sales tool that still accepted a weak development credential and had no rate limits. First access was Oct 25; bulk retrieval Oct 30. WorkOS and an independent forensics firm correlated the researcher's evidence with logs. The researcher says all obtained data was deleted and never shared. Exposed data remains CRM contacts, account histories, sales/support/Slack excerpts (some end-user emails), and product-usage metadata — not production customer app data.

[← Context & Harness Attacks](/context) · [markdown](/items/workos-nov-6-update-escaped-artex-agent-researcher-sandbox.md)

official (WorkOS update bulletin, Nov 6); researcher-attributed unintended ARTEX agent escape; independent forensics validated; data deleted per researcher

2026-11-06

### WorkOS Nov 6 update: the Oct 30 sales-tool access was an escaped ARTEX AI agent from a researcher's sandbox — not a malicious intrusion

- **Root-cause gaps WorkOS lists:** internet-exposed internal API; weak/dev credential in production; no rate limits; detection delayed five days.
- **IOCs (official):** IPs `154.64.252.77`, `183.1.65.106`, `95.169.20.166`, `203.190.1.60`; UAs `ArtEx/1.0`, `infra-hunt/1.0`, `artex-enrich/1.0`, `norma/0.4`, `verify-911-probe/1.0`; JA4 hashes in the bulletin.
- **Harness lesson:** agent sandboxes need hard egress allowlists — an 'escaped' eval agent looks identical to an attacker using the same framework. WorkOS now red-teams its own edge with ARTEX.
- Supersedes the earlier 'likely ARTEX' framing with a confirmed researcher attribution; phishing risk from the leaked CRM context remains.

Sources: [WorkOS: Update for October 30th, 2026 Data Security Incident (Nov 6)](https://workos.com/security/bulletins/update-for-october-30th-2026-data-security-incident) · [WorkOS: October 30th, 2026 Data Security Incident (Nov 2)](https://workos.com/security/bulletins/october-30th-2026-data-security-incident)
