Summary
- Google and Mandiant have observed offensive agentic frameworks used in live intrusion activity.
- The research also documents AI-assisted vulnerability analysis, abused coding-assistant configurations and stolen cloud AI resources.
- Google cautions that autonomous activity can create forensic noise and errors, limiting stealth and operational reliability.
Google Threat Intelligence Group and Mandiant have documented adversaries using agentic security tools during live cyber intrusions, with automated systems performing parts of reconnaissance, vulnerability analysis and attack execution. Their research, published on 16 November, describes a move beyond using language models merely to draft phishing messages or code. It also cautions that machine-driven activity can leave unusually visible traces and make mistakes that limit its usefulness to attackers.
The researchers report direct observations of certain open-source agentic penetration-testing frameworks, including CAIRN and PentAGI, being used in hostile activity. They also identify experiments with other tools and custom orchestration systems. These findings do not mean that every named framework has carried out a complete autonomous intrusion or that a fully independent agent has replaced the human operator in every campaign. The reported activity involves varying degrees of automation across specific tasks.
Agentic frameworks can link specialised tools to a system that selects subsequent actions based on the results of earlier work. In legitimate testing, those capabilities can support authorised security assessments. In hostile use, they can shorten the time between identifying exposed services and trying known attack techniques. Google’s account emphasises that the underlying weaknesses often remain familiar: compromised identities, unpatched services, permissive access and insecure integrations.
The report also describes automation around analysing software patches to identify weaknesses already fixed by vendors but not yet removed from every customer environment. That creates pressure on the period between publication of a security correction and deployment across production estates. Google distinguishes such use of disclosed flaws from the much harder task of reliably finding new zero-day vulnerabilities, and notes that language models also generate inaccurate exploit claims that require verification.
Another area of concern is the integration of coding assistants into enterprise software workflows. In several intrusions attributed to a suspected threat cluster called MIDNIGHT NEPTUNE, Google says attackers manipulated local AI assistant configuration files so that later developer activity could trigger malicious code. The research includes targeting of cryptocurrency organisations in North America and Europe. The precise attribution remains Google’s assessment rather than a judicial determination.
Such incidents illustrate how an assistant’s permissions and automatic execution settings can become part of a software supply chain. A configuration file may appear less sensitive than an executable, yet the tools that interpret it can call commands or access repositories. Attackers able to modify those settings may gain a route to subsequent execution by a legitimate user or process. The risk depends on actual permissions and tooling, not simply the presence of an AI model.
Google also reports financially motivated abuse of cloud AI accounts and infrastructure. Stolen access credentials can be used to consume model capacity and generate large unauthorised bills, while underground services offer access obtained through compromised accounts. The threat group’s capability in these cases depends heavily on traditional cloud identity compromise. The model adds a valuable resource for resale or abuse, but it does not remove the role of access management failures.
The researchers argue that autonomous offensive tools currently trade discretion for speed. Repeated requests, identifiable tool behaviour and mistakes can make a campaign easier to identify, while model-generated inaccuracies can undermine the operator’s objectives. Long-running espionage actors concerned about remaining undetected may therefore deploy such tools selectively, whereas criminals optimising reach may tolerate more obvious activity. Those are research assessments based on observed patterns, not certainties about all attackers.
Google observed AI-supported frameworks used for particular steps in live intrusions, while reporting limitations in autonomy and reliability. Google says it has disrupted abusive accounts and incorporated observations into defensive products. The material does not support treating all future intrusions as autonomous or assuming every enterprise AI integration is compromised.





