
Google says cybercriminals have reached a significant turning point in the use of Artificial Intelligence for Cyberattacks, with researchers uncovering a system capable of independently identifying software vulnerabilities, developing exploits and launching attacks with little or no human intervention. The discovery suggests that AI-powered cybercrime is moving beyond tools that simply assist hackers with individual tasks and towards automated systems capable of carrying out multiple stages of an intrusion.
Researchers from the Google Threat Intelligence Group (GTIG) and Mandiant describe the development as a significant threshold for Cybersecurity. Their investigation uncovered an exposed staging server hosting a framework called exploit_agent, which connected several advanced language models to automated vulnerability research and exploitation tools.
The system could analyse software changes, identify potential security weaknesses, create exploit code and execute attacks. Additional scripts handled activities such as internal network discovery, password guessing and webshell management. Together, these capabilities demonstrate how AI can coordinate an attack rather than merely provide suggestions to a human operator.
Google’s findings raise concerns about the speed at which attackers may compromise organisations. Security teams traditionally have time to investigate suspicious activity, identify exploited vulnerabilities and respond to intrusions. Autonomous systems could compress several of those steps into minutes, leaving defenders with less time to react.
Google Discovers an AI System That Can Find and Exploit Vulnerabilities
The most significant part of Google’s investigation was the discovery of exploit_agent, an automated system hosted on an unsecured staging server. Its architecture connected multiple cloud-based and locally operated language models with scripts designed to perform technical security operations.
Rather than relying on a person to manually investigate every potential vulnerability, the framework could examine software code and identify differences between versions released before and after a security patch. Those differences can reveal which parts of a program were changed to correct a weakness.
This approach is particularly important because patch comparisons can provide attackers with valuable information about newly disclosed vulnerabilities. Once a security update becomes public, criminals can study the changes to understand what was fixed and investigate whether unpatched systems remain exposed.
The discovered framework automated several stages of that process:
- Software analysis: It examined code before and after security patches to identify changes associated with vulnerabilities.
- Vulnerability identification: It searched for weaknesses that could potentially be exploited against vulnerable software.
- Exploit development: It used AI capabilities to generate code intended to take advantage of identified security flaws.
- Exploit execution: It could launch the resulting exploits automatically rather than waiting for an operator to initiate every action.
- Post-compromise operations: Additional scripts supported network discovery, password guessing and webshell management.
The combination is more concerning than any individual capability. AI-assisted vulnerability research is not entirely new, but connecting discovery, exploit creation and execution into a coordinated workflow reduces the amount of manual work required to conduct an intrusion.
What Happened After the Attackers Gained Access?
Google’s investigation found that the automation extended beyond the initial exploitation stage. Scripts on the exposed server were designed to map internal networks, attempt password guesses and install webshells.
A webshell is a malicious script placed on a compromised web server that can provide an attacker with a way to execute commands remotely. It can help criminals maintain access, explore a victim’s environment and carry out additional operations after the initial vulnerability has been exploited.
The framework also included functionality for deleting webshells installed by rival attackers. This suggests that the operator was not simply interested in gaining temporary access. The system contained mechanisms for managing compromised environments and dealing with competing malicious activity.
Google’s researchers also identified clues about the operator’s potential interests. Information recovered from the system suggested attention towards French media, entertainment and education websites. However, comments in the configuration file were written in Russian.
These details provide investigative context, but they do not independently establish the identity or nationality of the person operating the infrastructure. Language clues and target preferences can help researchers develop leads, but they are not sufficient on their own to attribute an attack to a particular country or organisation.
PaperCut Vulnerabilities Show How Quickly AI-Assisted Attacks Can Spread
Google’s findings also connect the growing use of automated attack systems with real-world exploitation of software vulnerabilities. One example involved two PaperCut print server flaws identified as CVE-2026-81578 and CVE-2026-82078.
According to the supplied investigation details, attackers exploited the vulnerabilities in August 2026. When chained together, the flaws allowed unauthenticated remote command execution, meaning an attacker could potentially run commands on a vulnerable server without first logging in through a legitimate account.
The timing of the subsequent activity highlighted the difficulty defenders face when newly disclosed vulnerabilities become attractive targets. Cybersecurity company GreyNoise reported a campaign that began four days after disclosure and targeted hundreds of organisations using the same weaknesses.
The suspected attacker appears to have used AI tools to help develop and test exploits and assemble a list of potential targets. After gaining an initial foothold, the operation reportedly expanded its access to domain administrator privileges within minutes.
Domain administrator access can give an attacker extensive control over an organisation’s Windows-based network environment. Depending on the configuration and security controls in place, this level of access may allow criminals to manage accounts, access sensitive resources and move between connected systems.
The important distinction is that AI does not necessarily need to invent an entirely new vulnerability to make an attack more dangerous. It can also accelerate the process of identifying exposed systems, adapting existing exploit techniques and carrying out actions after an initial compromise.
Google Identifies More AI-Powered Attack Frameworks
The exploit_agent discovery is part of a wider pattern that Google says it has been monitoring. Its researchers are tracking several open-source frameworks designed to automate offensive cybersecurity operations, including ARTEX, CAIRN, PentAGI, CyberStrikeAI and HexStrike AI.
These tools are designed to combine different security and attack capabilities into workflows that can perform tasks with varying levels of human involvement. Depending on their configuration, they may support reconnaissance, vulnerability scanning, exploit development and post-compromise operations.
Google has confirmed that CAIRN and PentAGI have been used against live targets, demonstrating that these frameworks are not limited to laboratory experiments or demonstrations.
From August through early October 2026, Google observed one threat actor operating a customised version of PentAGI to probe numerous organisations across Asia for security weaknesses. The activity illustrates how attackers can adapt publicly available automation frameworks to support campaigns against real organisations.
Google’s researchers described the current trade-off between speed and stealth in their assessment:
“While these frameworks are capable of significantly compressing attack timelines to machine speed and achieving operational objectives, they currently trade stealth for velocity.”
In other words, automated systems can perform operations quickly, but their behaviour may be noisy or easier for security products to detect. Rapid scanning, repeated tool calls and unusual network activity can create signals that defenders may identify.
However, Google expects that limitation to become less significant as these systems improve. The researchers anticipate reduced human involvement alongside more sophisticated and stealthy operations.
Zero-Day Exploits Are Becoming a Growing Challenge
Another concerning development highlighted by Google is the increasing volume of zero-day exploitation. A zero-day vulnerability is a software weakness that is exploited before the affected vendor has released an effective security patch or before defenders have had a meaningful opportunity to protect against it.
Google recorded at least eight zero-day exploits per month in 2025, compared with 11 per month in 2026. Activity peaked at 22 in August 2026, while the overall number of recorded exploits nearly doubled.
These figures point to a changing vulnerability landscape in which attackers are finding and exploiting weaknesses at a faster pace. AI could contribute by helping researchers and criminals analyse code, identify promising targets and develop proof-of-concept exploits more efficiently.
However, an increase in exploit activity should not automatically be attributed entirely to AI. Vulnerability exploitation has several drivers, including software complexity, security research, criminal activity and the availability of exploit-development tools. AI is an additional capability that can accelerate parts of the process.
Fake Exploits Can Also Overwhelm Security Teams
AI-generated security content creates another problem for defenders: not every exploit demonstration circulating online is genuine.
After PHP patches were released on October 21, 2026, hundreds of purported exploit demonstrations appeared. Many were reportedly broken and likely generated using AI tools.
Although these demonstrations did not necessarily represent working attacks, they created additional verification work for security professionals. Teams had to distinguish genuine vulnerabilities and usable exploit code from inaccurate or fabricated examples.
This creates an unusual situation in which AI can increase the workload of defenders even when the material produced by attackers is technically unreliable. Security teams must investigate more claims, validate more code and determine which threats require immediate action.
Criminals Are Also Stealing AI Computing Resources
Google’s report highlights another growing criminal activity involving artificial intelligence: LLMJacking. Instead of using AI systems only to attack other organisations, criminals are stealing access to legitimate AI infrastructure and reselling the computing resources.
Mandiant responded to several LLMJacking incidents during 2026 in which attackers drained victims’ cloud computing resources and language-model quotas. The resulting fraudulent usage bills reportedly exceeded $100,000 per day in some cases.
In the most severe incidents, losses reached approximately $1 million before the victims completed cleanup and remediation. The compromised access could also be resold through underground marketplaces known as LLM Mixers.
These marketplaces offered access to stolen AI services at advertised rates ranging from $0.10 to $25 per million tokens. APIMart reportedly applied a five-times surcharge when prompts exceeded 100,000 tokens.
The business model creates opportunities for criminals who want to use expensive AI services without paying legitimate providers. Stolen accounts and cloud credentials can give buyers access to computing capacity that would otherwise require verified accounts and paid subscriptions.
For businesses, the consequences can extend beyond an unexpected bill. Compromised AI credentials may expose sensitive workloads, create compliance problems and allow unauthorised users to access services connected to internal applications.
Automated AI Agents Are Being Used to Operate Fake Online Identities
Google also identified a contractor developing a platform capable of managing more than 100,000 accounts. The system reportedly combined virtual phones and automated agents to operate thousands of fake online personas.
Such infrastructure can support large-scale online manipulation by making it easier to manage numerous accounts without requiring a human operator to control each one individually. Depending on how it is deployed, similar systems could be used for spam, fraudulent engagement, impersonation or coordinated influence activity.
The scale of the platform is particularly relevant because fake accounts have traditionally required considerable time and operational effort to create and maintain. Virtual devices and automated agents can reduce that burden, allowing a relatively small group to manage a much larger network of online identities.
The discovery also illustrates that AI-enabled cybercrime is not limited to traditional malware or attacks against servers. Automated systems can target the wider digital environment, including online platforms, identities and account infrastructure.
Claude Code Configuration Files Become a Malware Delivery Route
Google’s researchers also reported activity involving cryptocurrency companies across two continents. Suspected MIDNIGHT NEPTUNE actors modified Claude Code configuration files so that launching the development tool would trigger the download of a backdoor.
This technique is particularly concerning because developer tools are often trusted components of everyday software development workflows. If an attacker can modify a configuration file that executes automatically when a tool starts, the developer may unknowingly trigger malicious activity while performing routine work.
A backdoor can provide attackers with a persistent route into a compromised system, potentially allowing them to execute commands, collect information or download additional malicious components.
The incident highlights the importance of protecting developer environments, configuration files and software supply chains. Organisations should not assume that a familiar development tool is safe simply because it is widely used or comes from a legitimate provider.
Google’s findings do not mean that Claude Code itself is malicious. The reported activity involved attackers abusing its configuration mechanism to deliver malware.
Why Fully Autonomous Cyberattacks Are Difficult to Defend Against
Traditional cyberattacks often involve several stages that require an attacker to make decisions, review results and manually execute commands. Even sophisticated criminal groups need time to analyse targets, adapt tools and respond to defensive measures.
Autonomous AI systems can reduce the time required for these activities by connecting multiple operations into a single workflow. A system might identify a vulnerable server, develop an exploit, test it, execute it and then begin exploring the compromised network.
This creates several challenges for defenders:
- Shorter response windows: Attacks that once required hours or days of manual work may progress through multiple stages in minutes.
- Greater attack volume: Automated systems can investigate more potential targets without requiring an operator to supervise every scan.
- Faster vulnerability exploitation: AI-assisted analysis can help criminals understand newly disclosed patches and identify systems that remain vulnerable.
- More complex investigations: Security teams may struggle to distinguish automated attacks from legitimate software activity, particularly when tools generate large volumes of requests.
- Lower barriers to entry: Attackers with limited technical expertise may be able to operate sophisticated frameworks by combining existing tools with language models.
Nevertheless, autonomous systems are not necessarily more effective in every situation. Google acknowledges that hallucinated information, unreliable exploit code and noisy tool behaviour can reduce their success rates.
These limitations may explain why some sophisticated espionage groups are expected to use AI selectively rather than handing entire operations over to automated systems. Human operators may still be better positioned to make sensitive decisions, evaluate uncertain intelligence and avoid actions that could expose an operation.
What Organisations Should Do to Prepare for AI-Driven Attacks
Google’s findings reinforce the need for organisations to treat vulnerability management and incident response as continuous security priorities. If attackers can move from vulnerability discovery to exploitation more quickly, defenders must reduce the time between identifying a weakness and applying an effective mitigation.
Organisations should prioritise the following measures:
- Patch internet-facing systems quickly: Publicly accessible servers, remote access tools and business applications should receive priority when critical vulnerabilities are disclosed.
- Monitor exposed infrastructure: Regularly identify internet-facing assets, outdated software and systems that have been forgotten or are no longer actively maintained.
- Strengthen identity security: Use multifactor authentication, strong access controls and privileged account monitoring to reduce the risk of attackers escalating access.
- Watch for unusual network activity: Monitor unexpected scanning, suspicious authentication attempts, lateral movement and unexplained outbound connections.
- Protect developer environments: Review configuration files, software dependencies and automated startup processes that could be abused to deliver malware.
- Secure AI credentials: Apply strict permissions to cloud and model-provider accounts, monitor token consumption and rotate credentials when compromise is suspected.
- Prepare automated response procedures: Security teams should establish clear processes for isolating affected systems, revoking compromised credentials and investigating suspicious activity.
Organisations should also test their defensive systems against automated attack techniques. Traditional security exercises may not fully represent the speed and scale of operations performed by AI-enabled frameworks.
Has AI Crossed a New Cybersecurity Threshold?
Google’s discovery suggests that AI-powered cyberattacks are entering a more advanced phase. The important development is not simply that criminals can ask a chatbot to write malicious code. It is that connected AI systems can increasingly perform a sequence of offensive security tasks, from identifying vulnerabilities to launching exploits and managing access.
The exploit_agent framework provides a concrete example of this transition. Its ability to analyse patches, generate exploits and automate post-compromise activities demonstrates how language models can become components of a broader attack infrastructure.
However, the findings do not mean that every cyberattack is now fully autonomous or that human hackers have become unnecessary. AI-generated exploits can still fail, automated tools can produce detectable activity and complex intrusions may require human judgement.
The more immediate concern is that the balance between attackers and defenders could shift as these systems become faster, more reliable and harder to detect. Criminals may need fewer people to conduct operations that previously required specialised technical teams, while defenders face an expanding volume of activity to investigate.
Google expects human involvement to decline as autonomous attack frameworks become more sophisticated. For security teams, that means preparing for a threat environment in which vulnerability discovery and exploitation may happen at machine speed, leaving less room for slow patching, manual investigation and delayed incident response.
For breaking news and live news updates, like us on Facebook or follow us on Twitter and Instagram. Read more on Latest Technology on thefoxdaily.com.

COMMENTS 0