Google identifies first instance of fully autonomous AI-driven cyberattacks in the wild

In a significant escalation of digital warfare, Google’s Threat Intelligence Group (GTIG) and its subsidiary Mandiant have confirmed the first observed instance of attackers utilizing artificial intelligence to autonomously identify, develop, and deploy software exploits without human intervention. This development marks a threshold moment in cybersecurity, signaling that the theoretical risks long debated by researchers have officially migrated into active, real-world weaponization.
The discovery centers on a sophisticated tool identified by Google researchers as "exploit_agent," which was uncovered on an exposed staging server. This framework represents a departure from traditional "human-in-the-loop" hacking, as it possesses the capability to scan target software, compare pre- and post-patch code, formulate a functional exploit, and execute an attack sequence independently.
The Mechanism of Autonomous Exploitation
The "exploit_agent" tool functions by leveraging a combination of frontier and locally hosted AI models to perform automated vulnerability research. By systematically analyzing the differences between patched and unpatched versions of popular web software, the agent can reverse-engineer the specific flaw that necessitated the security update. Once the vulnerability is identified, the agent writes the exploit code, tests it, and executes the breach—all within an automated pipeline.
This methodology effectively weaponizes the disclosure process. In the cybersecurity industry, the time between a patch release and a full public disclosure is often referred to as the "window of exposure." AI-driven agents drastically shorten this window, potentially giving defenders only hours—or even minutes—to secure their systems before autonomous attacks begin.
Chronology of the Threat Landscape
The evolution toward fully autonomous threats has been rapid. In the third quarter of this year, GTIG’s quarterly intelligence reports explicitly stated that while AI was being used for auxiliary tasks like phishing email generation or code obfuscation, no fully autonomous attack pipelines were being observed against external targets.
The shift occurred late this year, culminating in the discovery of the staging server. Evidence found on this server provided a detailed look at the attacker’s operational infrastructure. Beyond the AI exploit tools, the server contained a repository of stolen victim data, scripts for brute-force password guessing, and automated webshell deployment tools. Notably, the server also contained functionality to scan for and remove competing webshells—an indicator that the attacker was actively "cleaning up" compromised servers to maintain exclusive control over their victims.
Configuration files recovered from the site contained comments written in Russian, and victimology data indicated that the primary targets were French media, entertainment, and educational institutions. This suggests a targeted, coordinated campaign rather than indiscriminate "script-kiddie" activity.
The Patch-to-Exploit Pipeline
Google’s analysis suggests that patched vulnerabilities represent the "low-hanging fruit" for AI agents. By analyzing public release notes, code repositories, and security advisories, an AI can aggregate data faster than a human team.
A clear example of this phenomenon occurred in August, involving two vulnerabilities in PaperCut print servers. Within 24 hours of the public disclosure, security researchers—using AI tools to assist their defensive work—were able to create proof-of-concept exploits. However, by the fourth day, malicious actors had likely utilized similar AI-augmented workflows to initiate a campaign against hundreds of organizations. Data from the network security firm GreyNoise confirmed the rapid proliferation of scanning activity following that disclosure.
A similar trend was observed in October regarding PHP vulnerabilities. Following the release of security patches, hundreds of proof-of-concept exploits surfaced across public repositories. A significant portion of these were deemed to be low-quality or non-functional, leading GTIG to conclude that these were the products of generative AI tools that lacked the final verification step provided by human experts.
Proliferation of AI Hacking Frameworks
The threat is no longer limited to bespoke, private tools. GTIG reports that several distinct groups, including financially motivated cyber-criminal syndicates and state-sponsored actors suspected of ties to China, are integrating open-source AI hacking frameworks into their standard operational procedures.
One such framework, known as CAIRN, has been identified in live attacks against critical infrastructure, including the U.S. aerospace sector. While these current iterations of AI-driven frameworks are "noisy"—meaning they leave distinct, recognizable traces in system logs and network traffic—Google warns that this is a temporary state. As these models become more refined, they will inevitably become stealthier, requiring a fundamental shift in how security operations centers (SOCs) detect unauthorized activity.
Exploitation of AI and Cloud Credentials
The danger is compounded by the trend of attackers specifically targeting the infrastructure that runs AI models. The report highlights instances where threat actors compromise cloud and AI service credentials to run their own malicious models on the victim’s dime.
In some documented cases, the fraudulent computing costs incurred by the victim reached over $100,000 in a single day. A notable attack vector involved the LiteLLM gateway, where attackers exploited vulnerabilities to hijack underlying API tokens. In another instance, a suspected North Korean group gained unauthorized access to Claude Code settings, modifying them to ensure that every time an employee launched the application, a backdoor was automatically downloaded.
In response to these findings, Google has proactively disabled accounts and cloud assets identified as being involved in the misuse of its Gemini models, and has implemented tighter security controls on AI-related gateway services.
Analysis of Implications
The emergence of autonomous AI agents in the cyber-attack lifecycle represents an inflection point. Historically, the limiting factor in any cyberattack was the human component: the time required for a researcher to understand a vulnerability, write an exploit, and launch an attack. By removing this bottleneck, AI allows for an almost infinite increase in the scale and speed of offensive operations.
The implications for enterprise security are profound. Static defense models—which rely on patching cycles and manual threat hunting—are increasingly inadequate. Organizations must now transition toward "AI-resilient" architectures, which include:
- Automated Incident Response: Systems must be capable of identifying and isolating threats at machine speed, as human analysts will no longer be able to react in time.
- Behavioral Analytics: Since AI-driven attacks may not match known malware signatures, defenders must focus on identifying anomalous behaviors in system processes, such as unexpected code execution or unusual outbound traffic from staging servers.
- Hardened Supply Chains: As demonstrated by the exploitation of print servers and gateways, third-party software and API intermediaries have become high-value targets. Organizations must apply zero-trust principles to all integrated AI tools and software gateways.
Conclusion and Official Outlook
GTIG remains clear in its assessment: the use of AI in cyberattacks is not a passing trend but an evolving reality. "We believe it is highly likely that threat actors will continue to use these frameworks in campaigns targeting organizations globally," the report stated.
As the technology matures, the "noise" currently associated with these attacks will likely diminish, making them significantly harder to detect. For the security community, the message is one of urgency. The same tools being built to defend networks—automated vulnerability scanning, rapid patching, and predictive threat modeling—are being repurposed by adversaries to dismantle them. The race between offensive AI and defensive AI has officially moved from the laboratory to the front lines of global digital infrastructure.







