decrypted · 5 november 2026 · ransomware and cybercrime · vulnerabilities and patching · ai and llm security

A sales tool, an AI toolkit and the customer conversations it exposed

A company that sells login infrastructure to other software firms has told customers that an attacker, using an AI toolkit to do the legwork, reached an internal sales application and read customer conversations. WorkOS supplies SSO (single sign-on, one login for many apps) and user management to developers. The headline risk is not stolen code. It is a very convincing phishing email.

What was taken

WorkOS says it identified unauthorised access on 30 October to an internal application used by its sales team. It reports no evidence that application data or production services were touched. What was affected, it says, was CRM (customer relationship management) and support data: names, contact details, job titles, account histories, and summaries and excerpts from sales calls, support discussions and customer Slack conversations. It has not said how many customers are affected, or how the attacker got in.

The machine behind it

The bulletin describes an "AI-assisted offensive security framework" that automated reconnaissance and exploitation attempts. It does not name the tool, but BleepingComputer reports that WorkOS's founder named it as ARTEX AI, citing request signatures he posted, so treat that as the founder's account relayed by the press. Separately, BleepingComputer reports CrowdStrike's finding that a group it calls Vapor Panda used ARTEX with language models to run several intrusions at once against more than 200 organisations in 14 countries.

Picture a burglar who once tried one door at a time. He now sends a crowd of tireless apprentices to rattle every handle on the street and report back. There are simply far more attempts per hour, so the door nobody got round to fixing is found sooner.

What the defender sees

Noise, mostly. BleepingComputer's account of the Dutch volunteer group DIVD, breached by the same framework in September, describes more than 450,000 requests and 2,500 scripts, which DIVD called "super loud and messy", and a jump from first access to root in 12 minutes. WorkOS investigators reportedly spotted distinctive user agent strings, which only helps if you keep and read logs for internal tools.

The Secure by Design lesson

An internal sales tool is rarely treated as crown jewels, yet here it held fragments of customer conversations. Four design decisions would shrink the damage:

Keep the tool off the open internet, behind device-aware access, so the apprentices find no handle to rattle.
Store summaries, not transcripts, and delete them on a schedule. Data you no longer hold cannot be read.
Alert on request volume and odd user agents, because machine-speed attacks are rarely quiet.
Assume the contact data leaks, and brief staff and customers before it does.

For UK organisations, call summaries naming your staff are personal data, so ask your suppliers what their sales and support tools retain. If a vendor writes to you, follow the bulletin's own advice: verify unexpected requests, even ones that mention a real colleague or support conversation, before you act.

Also this week

Zyxel switches: a fix that did not hold. On 3 November Zyxel published an advisory for CVE-2026-107441, an out-of-bounds write in the web management code of its GS1900 series switches. The advisory says an unauthenticated attacker on the LAN (the network inside one building) could send a crafted HTTP request and possibly run operating system commands, and it lists fixed firmware for ten models. It does not mention exploitation. BleepingComputer, citing Binarly and GreyNoise, reports that an exploit worked against firmware patched in June for an earlier flaw, and that 996 devices in 48 countries were compromised; treat those figures as reported. The design lesson is that a check bolted onto one parameter, while the surrounding parser stays permissive, invites a second round. If you run these small switches in offices or branches, update them and keep their management pages off untrusted networks.

DreamHost: two months inside the file manager. The US hosting provider has told customers that between August 2026 and 23 October an unauthorised party exploited a vulnerability in the server behind its web-based File Manager and obtained the FTP and SFTP usernames and passwords of people who logged in through it. Session data for a limited number of webmail users may also have been exposed, and the notice says email contents and files may have been accessed. DreamHost says it found no evidence of access to control panel logins or billing details. It gives no count of affected customers. If you host with them, DreamHost's own advice is to check site files for unexpected changes, review email forwarding rules and change reused passwords. Password-only file transfer logins handled by a web tool are a standing prize; key-based access and short-lived sessions make a stolen login far less useful.

Sources

If your suppliers hold more of your conversations than you would like, get in touch.