Technology

Cyber attacks turn autonomous: Google reports AI agents now run intrusions at machine speed

Google's Threat Intelligence Group says attackers are using autonomous AI frameworks to find and exploit software flaws with little human involvement, including a tool it calls exploit_agent.

Cyber attacks turn autonomous: Google reports AI agents now run intrusions at machine speed
Photo: U.S. Marine Corps photo by Lance Cpl. Juan Diaz / Wikimedia Commons, Public domain (cropped)

Google's Threat Intelligence Group says cyber attacks are moving toward autonomous operation, with threat actors deploying AI agent frameworks that can find and exploit software weaknesses with little or no human involvement. In its quarterly AI Threat Tracker report, published on Monday, the group said the speed of such intrusions now marks a new challenge for defenders. The report was released as "cyber attacks" drew strong search interest globally.

A new kind of automated intrusion

According to the report, adversaries have gone beyond manual prompting of AI models. They are now deploying multi-agent frameworks that can carry out several stages of an intrusion in sequence, including reverse engineering software patches to identify the flaw that was fixed and then building an exploit for it. Google said these pipelines can scale exploitation against internet-facing servers.

Google's researchers said they had observed multiple threat actors using open-source agentic penetration-testing tools to automate specific phases of live intrusions. The tools named in the report include CAIRN, CyberStrikeAI, HexStrike AI and PentAGI, along with ARTEX, which Google described as already widely publicised. Google said it had independently seen CAIRN and PentAGI used in the wild, and CAIRN was linked to intrusion activity against targets including in the US aerospace sector.

The company stressed that the basic methods have not changed. Attackers still rely on compromised credentials and unpatched systems. What has changed is how quickly the frameworks can chain attack steps together and exploit known weaknesses.

A separately reported AI-driven attack

Separately, UPI reported that Google had detected, for the first time, an AI framework that autonomously identifies and exploits software vulnerabilities without human intervention. According to UPI, the framework, which it said Google named exploit_agent, was found on a server and searched popular web application code for flaws before drafting and running exploits. UPI reported that Google did not attribute the framework to any particular group, although it found Russian-language comments in its configuration file.

UPI also reported that the attacks resulted in stolen data, but Google did not disclose how many organisations were affected, what kind of information was taken or how much damage was done. According to the same report, there appeared to be a particular interest in French websites linked to media and entertainment, as well as the education sector. These details come from press coverage of the report and have not been independently confirmed.

Other misuse of AI tools

The report also describes AI use beyond intrusions. Financially motivated actors are reportedly stealing access to large language model credentials and cloud resources from compromised enterprise systems, then reselling access through underground proxy services. Google said state-linked groups and commercial influence-operation vendors routinely use language models to produce synthetic media, build localised personas and automate distribution across platforms.

Google also pointed to espionage groups using language models for target research and data analysis. It described a supply chain concern in which attackers manipulate the configuration of autonomous coding assistants so that backdoors are dropped during local execution. Google said it had observed evidence of a China-linked group developing its own automated attack framework for likely espionage use, though the report does not confirm attribution in every case.

Google's response and what comes next

Google said its response combines safeguards built into its models, takedowns of misuse projects and accounts, and defensive products that monitor and respond at machine speed. In a separate post, the company pointed to its Google AI Threat Defense offering, which it said combines its Gemini models, Wiz context, CodeMender automated patching and Mandiant expertise.

The report signals that defenders may need to patch faster and automate more of their own detection and response. For now, the findings describe a trend rather than a single confirmed large-scale campaign, and the full scale of the activity is not yet clear. Further detail is expected as researchers and affected organisations review the findings.

Dateline: not stated in the source reports.

Ԝ

"

SADDAM ADIL

Editor-in-chief

Saddam Adil is a journalist and digital media professional focused on delivering clear, factual, and engaging news.

Read next

The Chopal Brief

The day's most important stories, straight to your inbox.