Google says attackers now let AI agents run intrusions largely unsupervised

Google Threat Intelligence Group and its incident response arm Mandiant say attackers increasingly rely on AI agents that carry out intrusions with little direct human involvement. The finding comes from the latest AI Threat Tracker, as reported by e27.
What the latest tracker reports
In its latest AI Threat Tracker, published early this week, Google Threat Intelligence Group and Mandiant describe the adversarial use of AI as having reached a "critical inflection point". As e27 summarises the document, the human attacker is becoming less a hands-on operator and more a supervisor of automated systems.
According to Google, a growing share of the actual break-in work is now done by software agents that scan, exploit and adjust on their own.
From chatbot help to autonomous intrusions
Google's account is that attackers have moved past asking chatbots for help with phishing, the messages that impersonate someone to obtain passwords or money. They are now deploying autonomous, multi-agent frameworks that chain together entire intrusions.
This matters because it shifts which party carries out the sequence of steps in an attack: software now handles work that once required a person to decide and act at each stage.
What remains unclear
The account rests on a single report, published by e27 and drawn from Google's tracker. The material available does not detail specific attacks, affected organisations or the frameworks involved, nor how widespread the practice is, so the scale of the trend cannot be independently assessed from this reporting.
Open questions
- How widespread are these multi-agent intrusions in real incidents?
- Which specific attacks or frameworks does the tracker document?