Google’s Threat Intelligence Group has identified a system capable of finding software vulnerabilities, generating exploits and launching attacks without human intervention at each stage. Russian-language instructions found in the system suggest a possible Russian-speaking operator, but researchers have not attributed the activity to a specific hacking group or government.
Google has uncovered what its cybersecurity researchers describe as the first observed use of artificial intelligence agents to fully automate both vulnerability discovery and exploitation, a milestone in the use of AI for cyberattacks.
The findings, published on November 16, 2026, by Google’s Threat Intelligence Group (GTIG) and Mandiant, concern a multi-agent hacking framework called exploit_agent, discovered on an exposed server used to prepare cyber operations.
The system could continuously examine software security updates, identify weaknesses, generate attack code and execute exploits without requiring human approval during the process.
Investigators also discovered data stolen from victims and tools designed for further activity inside compromised networks.
Of particular interest, the server contained Russian-language configuration instructions, and technical records suggested that the operator had focused on French websites in the media, entertainment, and education sectors.
However, Google has not identified the hackers’ nationality, established their connection to the Russian government or named a responsible organisation. The Russian-language evidence indicates a possible linguistic connection rather than confirmed state attribution.
How the autonomous hacking system operated
The investigation was published in Google’s report, GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold, authored by Aurora Blum, Kelli Vanderlee, Jess Xia and David Mainor.
According to the researchers, the discovery represents a departure from earlier cases in which hackers used large language models (LLMs) primarily to assist with individual tasks.
Previous examples involved AI-generated scripts, help with vulnerability analysis or partially automated attack sequences.
The newly documented framework combines those functions into a continuous automated process.
Instead of requiring an operator to examine each vulnerability and decide how to exploit it, exploit_agent was configured to perform several stages independently.
How exploit_agent’s automated attack process worked
1. Monitor software updates. Compare the code before and after security patches.
2. Identify vulnerabilities. Use AI models to examine code changes and potential weaknesses.
3. Generate exploits. Produce attack code for the identified vulnerabilities.
4. Execute automatically. Attempt exploitation without a human approving each step.
Based on Google’s technical description of the framework. This illustrates its reported capabilities, not a reconstruction of a particular victim’s intrusion.
The central technique involved analysing differences between patched and unpatched versions of popular web applications.
Software vendors release patches to correct security flaws. But those updates can reveal which parts of the original code were vulnerable.
Attackers can examine the changes and attempt to develop exploits against systems that have not yet installed the update.
By automating this process, the framework could potentially reduce the time between a security patch becoming public and attackers attempting to exploit the underlying vulnerability.
Google found stolen data and tools for further intrusions
The researchers uncovered the framework on an exposed staging server.
Such infrastructure enables attackers to prepare tools, organise operations, and store information obtained from compromised targets.
Alongside exploit_agent, Google identified data exfiltrated from victims and scripts capable of conducting additional activities after a successful intrusion.
These included internal network reconnaissance, password-guessing attacks, installing web shells and removing web shells left by other attackers.
A web shell is a program placed on a compromised web server that can provide an attacker with continued remote access.
The ability to remove other attackers’ web shells is noteworthy because it suggests the tools were designed for more than simply identifying vulnerabilities. They also supported activity after attackers compromised a system.
Nevertheless, the public report does not establish that every identified script was executed against every victim.
Google’s evidence demonstrates the existence of a configured autonomous exploitation framework and stolen data associated with the server. It does not provide a complete public accounting of every intrusion carried out using the system.
Why investigators suspect Russian-speaking hackers
One of the clearest clues about the possible operators came from the framework’s configuration files.
Google discovered instructions written in Russian concerning the use of API keys for AI models. The comments referred specifically to automated exploitation and the exploit_agent framework.
The presence of Russian-language instructions suggests that the system was developed, configured or used by someone comfortable working in Russian.
But language alone cannot identify the nationality or institutional affiliation of a cybercriminal.
Russian is spoken in multiple countries, and individuals operating across different jurisdictions can develop or modify software.
The investigation therefore provides no basis for concluding that the operation was directed by Russia’s FSB, GRU or another government body.
Cybersecurity publication The Register, in a report by Jessica Lyons on November 16, similarly described the activity as involving a suspected Russian-speaking operator but did not name a specific state or hacking organisation.
This is an important distinction for cyberattack attribution: identifying a language used in technical files is not equivalent to proving who controlled an operation.
French media and education websites were of particular interest
Google reported that technical data recovered from the staging server indicated particular interest in websites associated with France’s media, entertainment and education sectors.
The discovery raises questions about the intended targets of the operation.
Media organisations may hold sensitive information about employees, sources and internal communications. Educational institutions can maintain substantial amounts of personal and research data.
Both sectors also operate public-facing web services that may be vulnerable if software updates are delayed.
However, the identities of the affected organisations were not disclosed in Google’s public report.
Nor did the researchers establish whether the targeting reflected financial motives, espionage objectives or another purpose.
The concentration on French websites is therefore a documented characteristic of the observed activity, not proof of a specific geopolitical operation against France.
Google identifies a broader shift toward AI-driven cybercrime
The exploit_agent discovery is part of a wider trend documented by GTIG and Mandiant.
The researchers report that cybercriminals and espionage-related groups are increasingly adopting multi-agent frameworks to automate stages of cyber intrusions.
Their report identifies several such frameworks, including CAIRN, CyberStrikeAI, HexStrike AI and PentAGI.
These systems combine language models with conventional cybersecurity tools, allowing them to perform reconnaissance, analyse vulnerabilities and adjust their actions based on results.
Google also observed experimentation by financially motivated hackers and groups suspected of conducting espionage connected to China.
The findings show that autonomous cyber capabilities are not confined to one nationality or category of threat actor.
The technology can potentially reduce the expertise and labour required for certain attacks, although sophisticated operations still depend on technical infrastructure, reliable access and effective exploitation methods.
Security patches are becoming a race against automated exploitation
One of the most significant implications concerns the time organisations have to respond after vulnerabilities become public.
Traditionally, security teams have attempted to install patches before attackers could develop and deploy reliable exploits.
AI-assisted analysis can accelerate the process of turning disclosed vulnerabilities into attack tools.
Google illustrated this development with incidents involving PaperCut print management software.
In late August 2026, attackers exploited two vulnerabilities in PaperCut servers. Google cited research indicating that AI tools helped researchers produce working exploits rapidly after disclosure.
The company also referenced GreyNoise reporting on a subsequent campaign in which attackers likely used LLMs to assist in developing and testing exploits, identifying targets and carrying out attacks against hundreds of organisations.
These examples are separate from the exploit_agent discovery, but they support the broader finding that AI can accelerate the exploitation of vulnerable systems.
Independent cybersecurity expert Craig Costello, a professor at Queensland University of Technology, discussed the implications with Leonard Bernardone of the Australian Computer Society’s Information Age on November 19.
Costello explained that organisations remain exposed between the publication of a security patch and its installation.
The development of automated exploitation could narrow that period further, placing additional pressure on security teams responsible for large numbers of internet-facing applications.
Are autonomous attacks becoming more difficult to detect?
Automation provides attackers with speed, but Google’s findings also identify limitations.
The researchers observed that autonomous frameworks can generate substantial network activity and leave recognisable technical traces.
Rapid execution does not necessarily produce discreet execution.
AI-generated exploits can also fail, requiring additional testing and creating misleading technical material for security researchers to evaluate.
Google’s report notes that numerous apparently functional exploit demonstrations generated with AI have proved unreliable.
This presents a challenge for both attackers and defenders.
Attackers must determine whether automatically generated code actually works. Defenders must distinguish credible threats from unsuccessful or misleading exploit claims.
Google nevertheless expects these systems to become more capable as language models and autonomous frameworks improve.
The potential danger lies in combining rapid analysis, continuous execution and increasingly effective adaptation to obstacles encountered during an intrusion.
What autonomous cyberattacks could mean for critical infrastructure
The discovery has implications beyond the sectors identified in this particular operation.
Critical infrastructure operators frequently depend on software applications, remote-access systems and connected services.
If attackers can automate vulnerability research and exploitation, they may be able to test more exposed systems in less time.
The potential consequences would depend on the systems affected and the safeguards protecting them.
For example, an intrusion into an administrative web application does not automatically provide access to industrial control systems. Compromising operational technology generally requires additional access, knowledge or vulnerabilities.
Nevertheless, organisations that manage energy, healthcare, transportation and financial services may face increased exposure where vulnerable internet-facing systems provide entry points into wider networks.
The Google investigation does not demonstrate that exploit_agent was used against power plants, hospitals or other critical infrastructure. Such scenarios represent possible future risks rather than confirmed outcomes of this campaign.
Analysis: The important breakthrough is automation, not the hackers’ language
The most consequential element of Google’s findings is the apparent integration of vulnerability research, exploit development and execution into a single automated workflow.
Cybercriminals have long relied on automated tools to scan networks and attack known weaknesses.
What distinguishes the newly documented system is its use of multiple AI models to analyse software changes and develop exploitation methods without a human supervising every stage.
This could alter the economics of cybercrime. Tasks that previously required significant time from specialists may become easier to reproduce and scale.
A small group could potentially examine more software products and attempt more attacks than would have been practical through manual analysis alone. However, three limits remain important.
First, Google describes the discovery as the first such case it has observed, not proof that no earlier autonomous attacks occurred.
Second, the system’s capabilities do not establish that every attack it attempted succeeded.
Third, the Russian-language material does not provide sufficient evidence to attribute the operation to the Russian state.
The findings are therefore most significant as evidence of an emerging technical capability rather than proof of a new state-sponsored campaign.
Conclusion
Google’s November 16 investigation documents a development in cybercrime: the operational use of AI agents to automate the discovery and exploitation of software vulnerabilities.
The exploit_agent framework was found alongside stolen victim data and tools for conducting further activity in compromised environments.
Technical records suggested particular interest in French media, entertainment and educational websites, while Russian-language instructions indicated a possible Russian-speaking operator.
The identities and affiliations of those responsible remain unconfirmed.
For organisations defending their networks, the central implication is that the period between vulnerability disclosure and attempted exploitation may continue to shorten.
The wider challenge is no longer simply whether attackers can use AI to assist cyber intrusions. Google’s findings indicate that they are beginning to delegate entire stages of the intrusion process to autonomous systems.
Sources and attribution
- Primary investigation: Google Threat Intelligence Group — Agentic Threat Activity Crosses a Significant Threshold
- Published November 16, 2026. Authors: Aurora Blum, Kelli Vanderlee, Jess Xia and David Mainor. The investigation provides the principal technical evidence concerning
exploit_agent, the stolen data, Russian-language instructions and targeting patterns. - Independent reporting: The Register — Google: AI Agent Turned a Security Patch Into a Working Exploit With No Human Help
- Jessica Lyons, November 16, 2026. Reports on the discovery and the limits of attributing the operation to a particular group. Information Age — Google Uncovers AI Hacks With No Human Input
- Leonard Bernardone, November 19, 2026. Includes comments from Google senior analyst Kelli Vanderlee and Queensland University of Technology professor Craig Costello on automated exploitation and the implications for defenders.
