Kaspersky uncovers intrusions across Latin America and Africa in which attackers used CyberStrikeAI to automate reconnaissance, steal credentials, move laterally and compromise privileged administrator accounts.
Cybercriminals are beginning to use artificial intelligence not merely to develop malicious tools or craft convincing phishing messages, but to coordinate and accelerate entire attack operations inside compromised enterprise networks. A new investigation by Kaspersky Threat Research has uncovered attacks on educational organisations in Latin America and Africa in which threat actors used an AI-powered penetration-testing framework to conduct reconnaissance, steal credentials, move across systems and collect sensitive information.
Observed between September and October 2026, the intrusions demonstrate how AI agents can help attackers adapt their next steps based on information gathered during an ongoing compromise. The activity also highlights the growing security implications for educational institutions, which often operate complex IT environments involving student records, research databases, administrative systems and networks shared by large numbers of users.
Kaspersky researchers traced the attacks to the use of CyberStrikeAI, an open-source penetration-testing framework that combines a command-and-control (C2) system with a large language model (LLM). The architecture allows a human operator to define objectives while the AI model selects tools, generates commands and analyses execution results. An implant installed on a compromised computer executes those commands and sends the results back to the attackers’ server.
This arrangement effectively gives attackers an AI-assisted operational layer through which they can coordinate activities across a victim’s network. Rather than manually issuing every instruction and interpreting each result, operators can use the model to help determine subsequent actions based on the environment they have accessed.
From manual intrusion to AI-assisted compromise
One reconstructed incident illustrates how quickly an intrusion can escalate once attackers introduce an AI-powered framework.
According to Kaspersky, the attackers initially operated manually through a compromised Microsoft SQL Server. They used this access to gather information about the environment and prepare tools for further activity. Following a pause of approximately two weeks, they deployed the CyberStrikeAI implant.
The subsequent operation progressed rapidly. Within three to four days of deploying the implant, attackers had compromised domain administrator accounts, giving them highly privileged access to the victim’s environment. The final sequence, involving movement between computers, credential collection and database export, took no more than an hour.
The timeline illustrates the operational advantage that AI-assisted coordination can offer threat actors. Once an attacker has established an initial foothold, automated analysis and command generation can help accelerate discovery, identify opportunities for privilege escalation and guide movement through connected systems.
However, Kaspersky’s findings also clarify an important technical distinction: the AI model did not run on the compromised computers. Instead, it operated on the attackers’ server. The implant on the victim’s machine functioned as a conventional remote-control tool, executing instructions and returning results without making direct calls to AI services.
This architecture means that the AI capability was used to support the attacker’s decision-making and orchestration rather than functioning as an autonomous AI application installed inside the victim’s network.
“AI changes how attackers coordinate intrusions and the speed at which they operate, but well-configured security tools such as Kaspersky EDR can still detect and help stop the techniques they use.”
Nikita Nazarov, Head of Threat Exploration, Kaspersky
Familiar techniques, faster execution
Despite the use of an AI agent, the underlying intrusion methods remained recognisable. Researchers identified activities including network discovery, credential dumping, lateral movement, persistence through malicious scheduled tasks, and the collection and exfiltration of data.
The attackers’ objective was not fundamentally different from that of conventional cyber intrusions. What changed was the way individual stages could be coordinated and adapted.
Nikita Nazarov, Head of Threat Exploration at Kaspersky, said: “The attackers used an AI agent based on CyberStrikeAI to carry out a full attack chain and adapt their next steps within compromised networks. However, the framework still relied on familiar techniques and procedures, including discovery, credential dumping, lateral movement, and collection and exfiltration. AI changes how attackers coordinate intrusions and the speed at which they operate, but well-configured security tools such as Kaspersky EDR can still detect and help stop the techniques they use.”
The observation is significant for enterprise security teams. While AI can make attack operations more efficient, it does not necessarily eliminate the behavioural indicators associated with established attack techniques. Suspicious access to credentials, unexpected administrative activity, unauthorised scheduled tasks and unusual communication between systems can still provide opportunities for detection.
For defenders, this reinforces the importance of monitoring what attackers do after gaining access, rather than relying exclusively on identifying malicious files or blocking known attack signatures.
How Kaspersky detected the activity
Kaspersky reported that its endpoint and network security technologies identified multiple stages of the intrusions.
Kaspersky Endpoint Detection and Response Expert detected attack behaviours including credential dumping and malicious scheduled tasks. Its network detection capabilities identified CyberStrikeAI command-and-control traffic and suspicious queries used to map victims’ domains.
The company identifies the observed implants through the detection names HEUR.Win64.CyberStrike.gen and HEUR.Linux.CyberStrike.gen.
These findings underline the importance of combining endpoint and network visibility. Endpoint telemetry can reveal suspicious commands, attempts to access credentials and mechanisms used to maintain persistence. Network monitoring can help identify communications with attacker-controlled infrastructure and uncover activity spanning multiple computers.
When these capabilities are correlated, security teams have a better opportunity to reconstruct an intrusion, understand its scope and respond before attackers complete their objectives.
Kaspersky also cautioned against drawing premature conclusions about attribution. Researchers found Chinese-language comments in some scripts, but these artefacts do not establish the attackers’ identity or location. The company could not attribute the activity to a known threat group or confirm whether the same operator controlled all the identified infrastructure.
Why educational institutions need stronger defences
Educational organisations are attractive targets because their technology environments often combine a wide range of users, applications and information assets. Universities, colleges and research institutions may maintain extensive databases, cloud services, learning platforms, administrative systems and research environments, each with different access requirements.
A compromised account with elevated privileges can provide an attacker with opportunities to move beyond the system initially accessed. If credential management, network segmentation and monitoring are inadequate, the compromise of one server or user account can develop into a wider organisational incident.
The Kaspersky investigation is therefore relevant beyond the education sector. Any organisation operating interconnected networks and holding sensitive information could face similar risks when attackers use AI to accelerate reconnaissance and intrusion activity.
For CIOs and CISOs, the immediate priority should be to strengthen visibility into identity activity, privileged access and communications between critical systems. Domain administrator credentials must receive particular protection through least-privilege access, multi-factor authentication, secure administration practices and continuous monitoring.
Organisations should also review how quickly they can detect suspicious lateral movement and isolate compromised endpoints. The speed of the final stages in the reconstructed incident demonstrates why periodic security reviews alone may not be sufficient. Detection and response capabilities need to operate continuously, with clearly defined escalation and containment procedures.
A multi-layered response to AI-assisted threats
Kaspersky recommends that organisations strengthen endpoint monitoring through an endpoint detection and response solution such as Kaspersky Next EDR Expert. Such capabilities can help investigate suspicious commands, credential access and persistence mechanisms, while supporting response actions on compromised hosts.
Endpoint protection should be complemented by network detection and response. Solutions such as Kaspersky Anti Targeted Attack can help identify suspicious network traffic and investigate activity that spans multiple systems.
Threat intelligence is another important component. Security analysts need access to current indicators, malicious infrastructure information and detection knowledge to support investigations. Kaspersky recommends its Threat Intelligence Portal for analysing suspicious files and infrastructure, alongside Hunt Hub for threat-hunting and detection information.
Organisations that cannot maintain continuous internal monitoring should also consider round-the-clock detection and response support. Kaspersky Managed Detection and Response provides monitoring, investigation and response assistance beyond normal working hours.
These recommendations point to a broader requirement: security teams must move beyond isolated alerts and develop the ability to correlate endpoint, identity and network activity into a coherent picture of an attack.
The next phase of cyber defence
The use of CyberStrikeAI in real-world intrusions marks another step in the evolution of AI-assisted cybercrime. It shows how open-source frameworks and large language models can help attackers coordinate established techniques, interpret information gathered from compromised environments and accelerate the progression of an intrusion.
Yet the findings also offer an important lesson: AI does not make conventional security fundamentals irrelevant. Strong identity controls, least-privilege access, endpoint visibility, network monitoring, threat intelligence and timely incident response remain central to defending enterprise environments.
For educational institutions and enterprises alike, the challenge is to recognise that attackers may now use AI to make familiar techniques faster and more adaptive. Security strategies must evolve accordingly, combining established defensive controls with continuous monitoring and the ability to respond at the speed of modern attacks.
