Google has warned that cybercriminals are deploying artificial intelligence systems capable of identifying software vulnerabilities and launching exploits without human intervention, marking a significant escalation in automated cyberattacks.The company's Threat Intelligence Group said its latest investigation uncovered a framework that independently examined software updates, identified weaknesses disclosed by patches and attempted to exploit affected systems. Google declared that the “era of autonomous attacks has arrived”, warning that defenders face increasingly compressed response times.
The findings, published on Monday, November 16, describe a shift from hackers using AI to assist individual tasks towards systems coordinating several stages of an intrusion. Google said its researchers had identified, for the first time, attackers using large language model frameworks to automate both vulnerability discovery and exploitation without a person directing each step.
Investigators discovered an exposed staging server containing a framework designated exploit_agent. It connected multiple advanced and locally operated AI models, continuously compared versions of popular web applications and generated attack code based on differences identified after software patches were issued.
The system was also configured to execute exploits automatically. Evidence recovered from the server included information taken from targeted victims and scripts supporting internal network reconnaissance, password attacks and the installation of malicious web-based tools, according to Google's analysis.
Victim-related records suggested that the operator was particularly interested in French websites belonging to media, entertainment and educational organisations. The findings establish the framework's operational capabilities, although the available evidence does not demonstrate that every attempted exploit succeeded.
Google identified a broader increase in malicious use of autonomous tools, including ARTEX, CAIRN and PentAGI, which can coordinate specialised agents to discover weaknesses, test access and adjust tactics after receiving feedback. Some were developed for legitimate penetration testing before being adopted by attackers.
Its investigators observed CAIRN in intrusions involving targets in the United States aerospace sector. Between late August and early October, another operator used a customised PentAGI-based platform to scan organisations across Asia, including telecommunications companies and systems integrators.
The researchers said criminal groups and suspected state-linked espionage operators were experimenting with such platforms. They cautioned, however, that the technology remains uneven: automated operations often generate conspicuous network traffic, leave identifiable code traces and make mistakes that can expose attackers to detection.
Separate investigations by Google and cybersecurity specialists have shown how quickly autonomous frameworks can progress once valid credentials are obtained. Research into ARTEX documented intrusions in which attackers moved from initial access to deeper systems within hours, sometimes considerably faster, although individual results varied.
A central concern is the shrinking interval between publication of a security fix and the development of usable attack code. By comparing patched and unpatched versions, AI systems can concentrate on previously undisclosed technical details revealed by the repair itself, potentially exposing organisations that have delayed installing updates.
Google also cited attacks against PaperCut print management servers as an illustration of accelerated exploitation. Security researchers produced working exploits using AI within a day of disclosure, while subsequent malicious activity suggested attackers had used similar assistance to target numerous organisations. That case was not established as a fully autonomous operation.
The company acknowledged that AI-generated exploits can be unreliable. Following the disclosure of two PHP vulnerabilities in October, numerous purported demonstrations circulated without functioning as claimed, complicating the work of security teams assessing genuine threats.
Beyond direct attacks, Google's investigation documented the theft of cloud credentials used to consume or resell access to commercial AI services. In some incidents examined by its Mandiant security division, fraudulent charges exceeded $100,000 in a day, while accumulated abuse reached $1 million before containment.
The threat assessment also described AI-assisted influence operations involving synthetic identities, automated publishing and multilingual political messaging. Google distinguished improvements in the volume and apparent quality of deceptive material from evidence that such campaigns actually changed public opinion.
Google emphasised that automation does not mean every intrusion is fully independent. Human operators still choose targets, supply credentials and set objectives in many observed attacks, while software agents carry out selected tasks at speeds that complicate intervention.
For defenders, Google highlighted automated monitoring, tighter access controls and faster remediation. Its AI Threat Defense offering combines Gemini models, Wiz risk assessment, CodeMender patching technology and Mandiant intelligence. Google said its security operations division had also introduced detection rules intended to identify activity associated with ARTEX.
Topics
Technology