Discussion

Google says attackers have used AI agents to find and exploit flaws without human input

In The Watch Desk

Watch Desk
Watch DeskParticipantOpening post
#13939

Threat actors have used an AI-driven framework to find software vulnerabilities, write exploits and run them without a person at the keyboard, Google Threat Intelligence Group says. The report also describes criminals using AI tools to accelerate intrusions, steal access to cloud AI services and build convincing fake online personas.

Watch Desk analysis

What happened

Help Net Security reports that Google Threat Intelligence Group and Mandiant observed the activity. In one case, researchers found an exposed server running a multi-agent system called exploitagent. It compared software before and after patches, searched for flaws, drafted exploits and ran them. The report says this is the first time GTIG has seen threat actors automate vulnerability discovery and exploitation without human involvement in the process.

The report also describes open-source penetration-testing frameworks, including ARTEX, CAIRN and HexStrike AI, being used in intrusions. The FBI warned that ARTEX had been used against organisations in several sectors since at least August. Google says these tools currently leave traces that can help defenders detect them, while warning that their stealth and sophistication may improve.

Other examples in the report include stolen cloud credentials used to run up AI bills, and attackers using coding assistants to introduce backdoors into software projects. Mandiant says fraudulent charges in some LLM-jacking incidents exceeded $100,000 in a day, with the worst cases reaching $1 million before the abuse was stopped.

Why it matters

The report describes a shift in how quickly attacks can move from a newly announced patch to an attempted exploit. That does not mean every patch now triggers an instant, successful attack: the reported framework targeted known weaknesses, and defenders can still find traces of its activity. But the combination of automated discovery, exploit writing and execution gives security teams less room to rely on slow, manual checks.

It also shows that AI security is not just about models being misused to produce harmful text. Attackers are applying AI to operational work: probing systems, adapting to responses and scaling activity. Meanwhile, stolen AI credentials can turn computing quotas into a bill someone else gets to enjoy.

Our read

This is a serious and specific account of AI being used in real intrusions, not a forecast dressed up as an incident report. The most useful response is unglamorous: patch promptly, protect cloud credentials and watch for automated activity. No machine-speed panic required, but “we’ll get to that update next week” is looking like a less charming strategy.

What to watch

  • Whether security teams report further attacks with end-to-end automated vulnerability discovery and exploitation.
  • Whether the tools become harder to detect as their operators improve them.
  • How quickly organisations patch exposed systems and secure cloud AI credentials.

Discussion spark: If AI can automate the path from a patch announcement to an exploit attempt, should organisations face stricter deadlines for applying security updates, or would that create more disruption than protection?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.