Cyberattacks are becoming more autonomous, according to Google’s threat intelligence team
GTIG describes offensive agents, stolen AI resources, and influence operations. In response, Google is presenting an automated defense platform.
On an exposed attack staging server, Google researchers discovered a tool configured to analyze software patches, identify the corresponding vulnerabilities, then produce and execute attacks without human intervention. The recovered data suggests a particular interest in French websites associated with media, entertainment, and education. The case appears in the latest Google Threat Intelligence Group report, which brings together observations from GTIG and Mandiant.
The investigation describes a progression in the offensive use of language models: operators are now connecting analysis, tool use, and adaptation to results in automated workflows. The findings concern tasks and sequences observed during intrusions. They do not mean that every campaign described operates without a human operator. Patches become raw material
The discovered tool, dubbed exploitagent, compares versions of web applications before and after a patch to understand what changed. That analysis can reveal the weakness the patch is intended to eliminate, then inform an attack against installations that remain vulnerable. The critical window falls between the release of a patch and its deployment by users.
This method targets vulnerabilities already known or patched by the vendor, often referred to as n-days. It differs from searching for previously unknown vulnerabilities, or zero-days. According to GTIG, models are particularly useful for analyzing a bounded collection of code versions, release notes, and security research.
A GreyNoise investigation cited in the report provides another example. During a campaign against PaperCut print management software, the company observed AI agents being used to develop, test, and deploy attacks. It identified at least 440 compromised installations linked to 395 named organizations across 48 countries. In that campaign, less than four hours elapsed between the creation of the workspace and the first code execution on a real victim’s system. Autonomy that leaves traces
The misuse of security testing tools is another focus of the report. Multi-agent systems divide tasks, use existing software, and adjust their actions based on the responses they receive. GTIG reports deployments of CAIRN and PentAGI in live operations, while distinguishing those observations from experimentation detected with other tools.
This acceleration still has limits. Automated execution produces recognizable traffic and technical traces, reducing stealth. Models can also invent details, target nonexistent resources, or produce nonfunctional exploit code. The report describes a proliferation of false proofs of concept that create additional verification work for security teams.
Existing protections can also block some attempts. In the PaperCut campaign investigated by GreyNoise, a web application firewall prevented an attack against at least one protected server. Compute and influence also become markets
The theft of AI resources, known as LLMJacking, involves using a compromised organization’s accounts and cloud infrastructure to consume its model services or computing capacity. Intermediaries then resell that access. In incidents handled by Mandiant, fraudulent charges sometimes exceeded $100,000 in a single day; the most severe cases reached $1 million in cumulative costs before remediation. These amounts describe individual incidents, rather than an average cost.
Influence operations also use models to create fictitious identities, adapt messages to a region, and automate distribution. GTIG describes commercial providers, state-backed actors, and networks primarily seeking advertising revenue. Better content quality and greater output do not, by themselves, demonstrate a stronger influence on audiences, the report notes. Automated defense as the proposed response
The publication accompanies the promotion of Google AI Threat Defense. The offering combines Gemini and other models with Wiz’s contextual analysis and risk prioritization, CodeMender’s remediation capabilities, and Mandiant’s threat intelligence. The provider aims to connect exposure discovery, code analysis, patch preparation, and continuous monitoring. The report’s findings therefore also serve to present its commercial response.
The proposed architecture distributes workloads between lightweight models for continuous checks and more advanced models for high-risk vulnerabilities. The offering’s webpage also provides for Mandiant to help teams organize the deployment of AI-generated patches.