AI Searches for Vulnerabilities and Launches Attacks without Human Intervention
Google reveals the first observed use of a multi-agent framework to automate vulnerability discovery and exploitation without human intervention at every stage.
Google has disclosed an artificial intelligence system called “exploit_agent” capable of finding and analysing software vulnerabilities, developing exploit code and automatically launching intrusion attempts. The discovery illustrates how AI’s role in cybercrime is shifting from technical assistance towards executing advanced stages of attacks.
In a report published on Monday, Google’s Threat Intelligence Group described it as the first observed case of a multi-agent framework being used to automate vulnerability discovery and exploitation without human intervention at every stage.
Researchers found an internet-exposed server hosting the framework, alongside configuration files, code and tools associated with intrusion operations. Its contents included instructions for connecting to multiple AI models and using them to analyse software and develop exploits.
The system compares software code before and after security updates, identifying developers’ changes and reconstructing the vulnerabilities they addressed. It can then generate code to test exploitation of those flaws.
The danger is that security updates intended to protect software can become a source of intelligence for attackers. Comparing two versions can reveal a flaw’s location and how it was fixed, while many systems remain exposed because organisations delay installing patches.
Google observed the system analysing access-control mechanisms in version 2.19.25 of the Spectra WordPress plugin. It did not, however, provide evidence that the system successfully compromised the plugin or exploited a specific vulnerability.
Google did not identify the operator or attribute the system to a particular state or hacking group. Russian-language comments in some configuration files are insufficient to establish who was responsible.
The discovery is not isolated. In recent months, cybersecurity companies and government bodies have observed the spread of offensive frameworks using AI agents to coordinate reconnaissance, vulnerability discovery, intrusions, and data collection.
On November 13, the FBI issued a warning about a tool called ARTEX, which uses external AI models to automate penetration-testing tasks. The Bureau said criminal groups had used it against financial, government, technology, healthcare and defence organisations.
Researchers identified more than 1,600 exposed operating interfaces for the tool online in early November. The warning also documented attacks that leaked personal data belonging to approximately 67,000 people at South Korean financial institutions. It described a supply-chain compromise involving ConnectWise Cloud Backup that enabled access to more than 2,250 downstream customer environments.
The FBI said some operations progressed from initial targeting to obtaining elevated privileges within an organisation’s systems in less than three hours.
In a separate investigation, Palo Alto Networks’ security research unit disclosed the use of another framework, CAIRN, in 25 operations targeting at least 19 organisations during August and September 2026.
The attackers used Chinese models, including DeepSeek and Zhipu GLM, to run the agents. According to the report, they gained unauthorised access to ten organisations’ systems. The investigation documented the theft of more than one million records in one case, alongside the creation of persistent administrator accounts in production environments.
One prominent operation targeted servers running PaperCut, software used to manage organisational printing. GreyNoise documented the compromise of 440 servers belonging to 395 organisations across 48 countries through two vulnerabilities allowing unauthenticated remote command execution.
A reconstruction of the operation showed attackers obtaining domain administrator privileges in one organisation within minutes. They also compromised eleven organisations in 26 seconds by running hundreds of connection sessions and agents simultaneously.
Separately, Amazon disclosed campaigns using AI tools to locate and steal access keys for language models. The keys were then reused to run agents that continued searching for additional credentials and computing resources.
These operations show that attackers can exploit victims’ resources to cover their tools’ operating costs rather than relying exclusively on their own accounts and services.
The spread of these frameworks raises concerns about a lower technical barrier to conducting attacks. Rather than requiring extensive programming and cybersecurity expertise, an operator can specify an objective and leave agents to examine systems, test intrusion paths, analyse results, and adjust tactics.
Increasing autonomy does not mean AI can compromise any system without human assistance, however. These tools still frequently require initial direction. Their attempts can produce invalid code or misleading results and do not guarantee successful exploitation.
Organisations face an additional challenge in the speed of execution. Operations that previously took days can now be automated, with simultaneous attempts against numerous targets reducing the time available for security teams to detect and contain intrusions.
Technology companies, meanwhile, have begun deploying AI agents defensively to discover vulnerabilities, analyse code and fix errors before attackers exploit them.
The confrontation is therefore becoming a race between offensive systems learning to find and exploit weaknesses and defensive tools trying to identify and close them just as quickly. The danger does not necessarily lie in AI inventing new hacking methods, but in its ability to execute known techniques faster, at greater scale and with less human intervention.









