Google warns AI agents now run cyberattacks with no human in the loop

Google says attackers have crossed a threshold: artificial intelligence agents are now finding and exploiting software flaws on their own, with no human at the keyboard. The warning comes in a report from Google's Threat Intelligence Group (GTIG) and Mandiant, published on 16 November 2026.
The report, titled "Agentic Threat Activity Crosses a Significant Threshold", says it is the first time Google has observed threat actors fully automating vulnerability discovery and exploitation. Researchers found a multi-agent framework they call "exploit_agent" on an exposed staging server. It compares software code before and after a vendor publishes a patch, works out what the fix reveals, drafts an exploit and runs it, all without a person involved.
Google said the operator behind that framework appeared to focus on French websites in media, entertainment and education. In one case the framework analysed an access-control check in the Spectra WordPress plugin, version 2.19.25, also known as Ultimate Addons for Gutenberg.
The frameworks behind the shift
The report is not about a single tool. Google says adversaries are now deploying open-source, multi-agent offensive frameworks such as ARTEX, CAIRN, CyberStrikeAI and PentAGI to carry out multi-stage intrusions faster. GTIG observed CAIRN and PentAGI being used in real attacks, while CyberStrikeAI and HexStrike AI were still being tested.
CAIRN, which runs specialised agents through an observe, orient, decide and act loop, was used in intrusions against targets including organisations in the United States aerospace sector. A custom platform built on PentAGI scanned a wide range of organisations across Asia between late August and early October 2026. ARTEX, an autonomous penetration-testing framework pulled from GitHub in October, had already been linked to automated attacks on banks in Southeast Asia.
Google quoted its own findings bluntly: "Adversary tradecraft has continued to transition beyond manual prompting to deploying autonomous multi-agent offensive frameworks — such as ARTEX, CAIRN, CyberStrikeAI, and PentAGI — to execute multi-stage intrusions at greater operational velocity." Threat actors are also using large language models to automate the reverse engineering of patches, letting them rapidly weaponise known vulnerabilities and scale exploitation against internet-facing servers without human intervention.
Exploits nearly doubled in eight months
GTIG's data shows overall vulnerability disclosures and exploits nearly doubled between January and August 2026. Zero-day exploits averaged 11 a month and spiked to 22 in August. Google did not name any group or state behind the activity it observed.
The report also describes LLMjacking, in which attackers steal cloud credentials and computing resources to run or resell access to AI models. GTIG saw prices from $0.10 to $25 per million tokens on Russian- and Chinese-language markets. Google's SecOps team has released behaviour-based detection rules for ARTEX, and Mandiant has published a guide to help organisations secure their systems.
Google expects the trend to accelerate. It warns that sophisticated, fully autonomous agentic attacks could emerge within the next three to six months, and that as the frameworks mature they will involve less human input and become harder to detect. For defenders, the message is that patching speed now matters more than ever: once a fix is published, an agent can turn it into a working exploit in hours.
Source: Infosecurity Magazine, with further detail from Security Affairs.