Google published an analysis on November 17 examining the use of AI agents in cyber operations, with examples identified by Google Threat Intelligence Group and Mandiant. The report describes the automation of attack stages, multiple agents working within the same workflow and the theft of access to inference resources. Its observations come from investigations and activity tracked during 2026, rather than exclusively from incidents on the publication date.
The central development is the move from an assistant answering questions to systems executing successive tasks through tools. In the documented cases, these systems can analyze technical information, prepare actions and pass results to the next stage. Google presents observed activity and its own assessments; the report does not state that every recent attack is fully autonomous.

How agents change the progression of attacks
Google distinguishes observed use of CAIRN and PentAGI in intrusions from experimentation with CyberStrikeAI and HexStrike AI. For PentAGI, the company describes reconnaissance and scanning against organizations in Asia, tracked between late August and early October 2026. The documented observation period therefore precedes publication of the November report and identifies when that particular activity was tracked.
Google also tracks workflows in which multiple agents divide analysis and execution activities. One agent can receive a bounded task, with its result becoming an input for another. Automating coordination reduces manual interventions between stages, but system access, permissions and available tools continue to determine which actions can actually be performed by the system.
The mechanisms described include analyzing differences between software versions after a fix is published. Such a comparison can help identify the vulnerable component and prepare exploitation of systems that have not received the update. In this context, the publication of a patch and its actual installation in the application being used are two distinct events.
Vulnerabilities and compromised access remain entry points
The report does not replace the technical causes of incidents with a generic explanation about AI. Known vulnerabilities, outdated applications and access obtained through compromised credentials still feature in the scenarios analyzed. Agents can organize or accelerate activities built around these entry points without the model itself creating the initial vulnerability in the affected system.
One concrete example is a framework called exploit_agent, identified by GTIG on an exposed staging server. According to Google, it analyzed differences between web-application versions and automated exploitation stages. Telemetry found there indicated interest in French media, entertainment and education websites. The report connects this observation to the investigated infrastructure and the data available on that server.
The connection with web application security is direct. Updates, access control and configuration protection concern the very components an automated operation may encounter. Google's report adds examples of how they are used in attacks, without treating every website incident as an attack carried out with AI agents or assigning the same mechanism to unrelated incidents.
Four distinct mechanisms documented by Google
| Mechanism | Target | Essential distinction |
|---|---|---|
| Agent coordination | Stages of an operation | The degree of automation differs between cases |
| Patch analysis | Vulnerable software that remains outdated | A published fix also needs deployment |
| LLMJacking | Access to and consumption of AI resources | Resources are used through the compromised account |
| Configuration tampering | Coding-agent behavior | Project files can influence execution |
LLMJacking consumes AI resources through someone else's account
A separate part of the analysis concerns LLMJacking: unauthorized use of an organization's access to AI models and infrastructure. In this type of incident, the attacker seeks the ability to use the service through compromised credentials. The targeted resource is inference capacity associated with the account, with consumption and costs recorded within the account holder's infrastructure.
This mechanism differs from stealing a model's files. An attacker can abuse a service without downloading model parameters or administering the provider's infrastructure. Identifying the AI product used by a business therefore tells only part of the story: the account, access keys, limits and granted permissions define the access actually available through that implementation.
Mandiant reports that fraudulent charges exceeded $100,000 in a single day in some LLMJacking incidents investigated during 2026. In the most severe cases, cumulative resource abuse reached $1 million before remediation. These amounts describe cases reported by the company, involving different durations and resources; they are not an average cost assigned to an incident.
Coding-agent configuration becomes relevant to security
Google also describes tampering with Claude Code configurations in internal code repositories after initial compromise of cryptocurrency organizations in North America and Europe. The changes could trigger a backdoor download when the tool started. The observation concerns configuration integrity within the affected organization, rather than a general compromise of the service operated by the model provider.
A coding agent may have access to files, commands or services, depending on its implementation. The instruction interpreted by the model and the technical permission are separate layers, however. Changing the text guiding an agent does not automatically create a new operating-system permission, but it can influence how the agent uses the rights it already has.
This issue also concerns AI integration in custom applications. An integration includes the model, data, tools and rules governing the execution of actions. For a business connecting an assistant to documents or internal operations, the identity and permissions of the component actually executing the action remain concrete elements of the application's architecture.
What the evidence in the report represents
Google and Mandiant draw on incident-response information, tracked activity and platform defense mechanisms. Each category of evidence has its own context. An observation about an investigated attack does not automatically describe the frequency of the same mechanism across every economic sector or its worldwide success rate against organizations using similar software and services.
The full technical report is dated November 17, while Google's short overview appeared on November 16. These dates describe the publication of the materials, rather than a shared starting time for the campaigns analyzed. Preserving that chronology separates the new analysis from the age of the incidents the company has documented through its investigations.
For businesses using AI services, the new information is the documentation of concrete abuse mechanisms alongside established application-security problems. Models, agents, cloud accounts and connected software appear together in these investigations. The technical responsibilities remain identifiable: who has access, what they can execute and which resources they can use on behalf of the organization holding the account.