CyberAv3ngers: Iran-linked cyber research activity
OpenAI ● Covered by 4 sources
OpenAI shut down accounts linked to Iran's CyberAv3ngers hacking group. They were using ChatGPT to scope out industrial control systems and default passwords for critical infrastructure.
Based on reporting by OpenAI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI's latest threat report names a familiar villain: CyberAv3ngers, the Iran-linked group that U.S. and Israeli officials have tied back to the IRGC. This isn't the group's first rodeo. Back in 2023, CyberAv3ngers took credit for breaking into Unitronics programmable logic controllers at water utilities across the U.S. and Ireland, some of them running on nothing more than a factory-default password. OpenAI says accounts tied to the group were using its models to do reconnaissance work that looked a lot like prep for more of the same.
The activity OpenAI flagged wasn't flashy. No jailbreaks, no malware-writing sessions. Instead, the accounts were asking about industrial control system protocols, hunting for default logins tied to specific hardware and software vendors, and building out lists of potential targets, the kind of grunt work that used to eat up hours of manual searching. That's the pattern OpenAI keeps surfacing across its threat reports this year: state-linked actors treating AI less like a weapon and more like a very fast intern who never gets bored doing OSINT.
OpenAI's response was to ban the accounts and share findings with the usual circle of industry partners and government contacts. It's a familiar playbook by now, and one the company has leaned on repeatedly as it tries to prove it can police misuse without slowing down access for everyone else. Whether banning accounts actually stops a state-backed group with presumably endless burner emails is another question entirely, but it does at least put a marker down and generate a paper trail for defenders watching the same infrastructure.
What makes this case worth paying attention to isn't the sophistication, there wasn't much, but the target list. Water treatment plants and industrial control systems remain some of the softest infrastructure targets in the West, running decades-old equipment with credentials nobody bothered to change. AI didn't create that vulnerability. It just made it faster to find.
My take — AI-written commentary, not fact-checked reporting
None of this required a jailbroken model or some exotic prompt trick, just a chatbot answering questions any search engine could eventually surface, only faster and with less friction. That's the uncomfortable truth the AI safety debate keeps dodging: the real risk isn't models suddenly inventing zero-days, it's models making tedious reconnaissance cheap for groups who'd have done it anyway. Banning accounts is fine PR, but the actual fix is patching decade-old PLCs and retiring default passwords, something no amount of model alignment will solve for you.
Read more about this at: OpenAI
