Google has documented what it assesses to be the first AI cyberattack to automate vulnerability discovery, exploitation and post-compromise activity without a human in the loop. The operation used multiple AI agents to move from analysing security patches to compromising targets and stealing data.
The findings were published on 16 November 2026 by Google Threat Intelligence Group, working with Mandiant. Russian-language elements appeared in the attackers’ tooling, although Google has not attributed the activity to a named group, individual or state.
How Google discovered the autonomous AI cyberattack
Investigators found an exposed, attacker-controlled staging server containing a custom multi-agent framework called exploit_agent. The server provided Google and Mandiant with visibility into the system’s architecture, supporting tools and apparent targets.
Exploit_agent chained together several specialised AI agents backed by a combination of frontier and locally hosted language models. Instead of relying on a person to direct each stage, the system could continuously examine software changes, infer vulnerabilities, develop exploits and execute them against candidate targets.
Google described the discovery as a significant threshold in malicious AI use. Criminal and state-linked actors have previously used generative AI to write code, compose phishing messages, translate content and summarise stolen information. In those cases, people generally remained responsible for choosing targets and directing the intrusion.
This AI cyberattack was different because the pipeline operated across the vulnerability research, exploit development and attack stages without direct human control. Google says the activity progressed beyond experimentation and resulted in successful exploitation, post-compromise actions and the theft of victim data.
Evidence pointing to Russian-speaking operators
Russian-language comments and instructions were present within the operation’s code and configuration files. One configuration instruction reportedly told the user to place application programming interface keys for automatic exploitation in the file rather than entering them through a chat interface.
These clues support an assessment that the operators were Russian-speaking, but language alone does not establish identity or government sponsorship. Google therefore stopped short of connecting the AI cyberattack to a particular threat group or country.
How exploit_agent automated vulnerability exploitation
The central technique involved comparing source code before and after a vendor released a security update. This process, often called patch diffing, helps attackers identify the code changed to correct a vulnerability.
Once exploit_agent detected a relevant change, its AI agents analysed what the patch revealed about the original weakness. The system then proposed exploit code, refined it and attempted to execute the resulting attack against internet-facing applications.
The automated workflow covered several connected stages:
- Monitoring popular web application software for newly published security updates.
- Comparing vulnerable and corrected code to identify security-relevant changes.
- Inferring the underlying vulnerability and conditions required to trigger it.
- Generating and refining exploit code through multiple language models.
- Executing the exploit against possible internet-facing targets.
- Conducting reconnaissance and other activity after gaining access.
Combining these tasks matters because it removes pauses that usually occur when research is handed between people or teams. A successful system can also repeat the process across many applications and targets, allowing one operator to pursue a much larger number of opportunities.
Tools found for post-compromise activity
The staging server contained more than vulnerability research components. Google and Mandiant also identified scripts for internal reconnaissance, password brute forcing and webshell installation.
A webshell is a malicious script placed on a compromised server to provide persistent remote access. One recovered tool was designed to remove webshells installed by competing attackers, potentially helping the operators retain exclusive control of an affected website.
Together, these components show that the AI cyberattack was designed as a broader intrusion process rather than a stand-alone exploit generator. It could move from finding an initial weakness to establishing access and exploring the compromised environment.
Victims, affected software and exploitation status
Artefacts and victim information recovered from the staging server indicated interest in French websites across the media, entertainment and education sectors. Google has not disclosed the affected organisations or the total number of successful compromises.
The researchers described exploit_agent as targeting popular web application software, but they did not name the products, CVEs or software versions involved. No specific IP addresses, domains, file hashes or file paths were published for the staging server or its tools.
That lack of public indicators limits organisations’ ability to search for this particular operation using simple blocklists. It also means reports should not present any individual web platform or software vendor as the confirmed target of exploit_agent.
Google did confirm that the pipeline reached working exploit execution and subsequent intrusion activity. As of 16 November 2026, the company assessed it as the first observed operation to conduct fully automated vulnerability discovery and exploitation without a person directing the process.
The separate PHP patch-to-exploit example
Google also highlighted the speed of AI-assisted exploit development through a separate incident involving two critical PHP vulnerabilities, CVE-2026-109618 and CVE-2026-109619. Both were fixed on 21 October 2026.
Hundreds of purported proofs of concept appeared within days of the patches. Many were non-functional and assessed as likely generated with language models, but their rapid publication illustrates how quickly defenders can face both credible exploits and large volumes of unreliable attack code.
Google did not identify these PHP vulnerabilities as exploit_agent targets. They are relevant because they demonstrate the same patch-to-exploit pressure that an autonomous AI cyberattack could intensify.
Agentic attack frameworks are already in use
The report also described adversary adoption of open-source offensive frameworks including CAIRN and PentAGI. Google observed such agentic systems in activity involving an aerospace organisation in the United States and systematic targeting across organisations in Asia.
These frameworks can coordinate separate agents for reconnaissance, vulnerability assessment and exploitation. Google also cited a China-nexus actor developing an automated attack framework for likely espionage activity, although this was separate from exploit_agent.
The combined findings indicate that autonomous attack tooling is moving into real intrusions. Not every AI-generated exploit works, but a framework can test large numbers of options quickly and retain successful techniques for future campaigns.
Why the AI cyberattack matters to UK organisations
The immediate concern is the shrinking interval between patch publication and exploitation. Internet-facing applications can be scanned continuously, while automated systems analyse the vendor’s fix and begin testing possible exploits before many organisations have completed patching.
UK small and medium-sized organisations running public websites may be particularly exposed where patching depends on a hosting provider, external developer or scheduled maintenance window. Common PHP-based applications are relevant to this risk, although they were not confirmed as targets in the exploit_agent case.
Actions organisations should take now
Defensive priorities should focus on reducing exposure during the patch-to-exploit window. Organisations should:
- Identify internet-facing applications and confirm their software versions, owners and patch responsibilities.
- Prioritise urgent updates for externally accessible web software, including supported PHP installations and dependencies.
- Review web server logs for unexpected file creation, password guessing, unusual requests and signs of webshell activity.
- Test web application firewalls and monitoring controls against current exploitation techniques rather than relying only on published indicators.
- Maintain recoverable backups and a response process for taking a compromised website offline safely.
The exploit_agent discovery does not mean every cyberattack is now autonomous. It does show that defenders can no longer assume exploit development will be constrained by an attacker’s available time or specialist expertise.
Originally reported by cyberscoop.com.







