CyberStrikeAI Used in Real Attacks: Foothold in Two Minutes, Domain Admin in Three Days
Kaspersky has documented real attacks in which the open-source-based AI tool CyberStrikeAI was used against educational institutions in Africa and Latin America: a human hacker spent one day on an SQL server, and two weeks later an AI agent took over — a foothold in two minutes, stolen domain admin accounts in three to four days, and data exfiltration in under an hour. The key point: the attack left no AI traces on the victim machines.
The News
Kaspersky's threat intelligence team (CTI) has published an analysis of cases in which the open-source tool CyberStrikeAI — an AI-driven penetration testing tool — was misused in actual cyberattacks against educational institutions in Africa and Latin America (BigGo Finance).
The finding lands at the same time as Google Threat Intelligence Group's AI Threat Tracker, published on November 16, lists CyberStrikeAI among the autonomous offensive frameworks threat actors are adopting (Google Cloud).
Kaspersky has not named the affected institutions or countries beyond the regional description, and none of the victims have been independently confirmed. All the case details below come from Kaspersky's analysis, reproduced via BigGo Finance as a secondary source.
The Attack Sequence, Step by Step
Kaspersky's reconstruction divides the attack into two phases — a human one and an agent-driven one — with a marked pause in between:
- The SQL server, roughly one day. A human hacker first took control of the victim organization's SQL server to gather basic information and map the network structure. This took about one day.
- A pause of around two weeks. There was then a gap of roughly two weeks before the next phase. Why the pause occurred is not explained in the analysis.
- The AI agent is deployed. On the attacker's own C2 server, an AI agent was set up to drive the attack forward.
- A foothold in two minutes. After the agent was activated, the attackers secured a persistent foothold on the victim's network in just two minutes.
- Domain administrator in three to four days. Domain admin accounts were stolen within three to four days.
- Lateral movement and exfiltration in under an hour. The final phase — moving through the network and exfiltrating data — took less than one hour.
The timeline may be the most significant aspect of the case. The intrusion techniques themselves are not revolutionary; it is the tempo and persistence that distinguish this from a human-driven attack.
How It Worked: The Human Set the Goal, the LLM Ran the Details
Kaspersky describes a division-of-labor pattern: a human defined the goal, and a large language model (LLM) handled the detailed execution — that is, which commands to run next. The analysis concludes that the AI did not invent any new hacking techniques. It chained together known, established methods, but did so faster and more persistently than a human operator typically manages.
That means the attack's "intelligence" lay in the orchestration, not in the ingenuity. Each individual command could be recognized as familiar tradecraft; what can sometimes go unrecognized is the pattern over time when an agent iterates without tiring or losing focus.
Why the Victim Machines Had No AI Traces
One of the most consequential details in Kaspersky's reconstruction: the LLM making the decisions during the attack did not run on the victim machines, but on the attacker's own C2 server. The victim machines therefore showed no logs of connections to external AI services such as ChatGPT or Claude, and no resource-usage traces of the kind that arise when AI models run locally.
For the defense side, this means the most intuitive detection heuristic — "look for traffic to AI services or unusual GPU usage" — does not catch this case. The distinction between attacks where the model runs on the attacker's side and attacks with local inference thus becomes concrete and practical, not theoretical.
Who Was Behind It? Nobody Knows
Kaspersky found Chinese-language comments in the attacker's scripts, but could not identify the attacker's identity or location from that alone. The company also could not tie the activity to a known threat group, or confirm whether a single operator controlled all of the identified infrastructure. It is therefore unknown whether the attacks in Africa and Latin America were one undertaking or several independent ones.
The language of script comments is in itself a weak indicator — such material is easy to plant, and Kaspersky itself draws no geographic conclusion. Any connection to China or Chinese actors therefore cannot be documented.
The Broader Trend — and an Unresolved Tension
Google's Threat Intelligence Group writes in its AI Threat Tracker of November 16 that threat creators' techniques "have continued to move beyond manual prompting and toward the deployment of autonomous multi-agent offensive frameworks — like ARTEX, CAIRN, CyberStrikeAI, and PentAGI — to execute multi-step intrusions at greater operational speed" (Google Cloud).
It is worth noting a tension between the two sources: GTIG characterizes CyberStrikeAI primarily as something threat actors are experimenting with, while Kaspersky documents use in real, completed attacks. How these assessments should be reconciled is not clarified in the available material. In any case, the GTIG report does not document the same attack Kaspersky describes — it is a broader trend observation.
As a related trend, but clearly separate from the Kaspersky case: Halcyon's analysis, as reported by WIRED, estimates that around 111 ransomware groups claimed attacks between July and September this year, in what the researchers describe as an AI-fueled surge in ransomware activity (WIRED). This concerns a different attack pattern and cannot be used as documentation for the CyberStrikeAI case.
What This Means for Defense
Kaspersky's conclusion is direct: the defensive response should shift from malware detection to behavioral analysis. The company assesses that existing behavioral analysis rules could also have detected this attack — precisely because the attack used no new malware or new techniques, but known methods chained together in a recognizable behavioral pattern: suspicious access to an SQL server, abnormal authentication activity targeting domain admin accounts, and then massive lateral movement and data exfiltration.
It is a demanding but concrete formulation: what will expose agent-driven attacks is not signatures of malicious code, but detection of behavior that deviates from the norm in the network — regardless of how "intelligent" the execution is in origin.
Hong Jun-ho, a professor at the Department of Convergent Security Engineering at Sungshin Women's University, puts the risk sharply: "The biggest risk of AI agents is not necessarily that they create new hacking techniques" (BigGo Finance). The core of the risk, according to him, is speed and persistence — known techniques being executed far faster and more persistently than a human can manage.
Open Questions
Several key questions remain unanswered in the available material:
- Who were the victims? Neither the institutions nor the countries beyond "Africa and Latin America" are named, and no victims have publicly confirmed the attacks.
- One operator or several? Kaspersky could not determine whether a single actor controlled all of the infrastructure.
- Where the maturity assessments diverge: GTIG describes experimentation; Kaspersky describes completed attacks. The difference may be due to different source material, different definitions, or different cases, but this is not clarified.
- The two-week pause: What happened in the gap between the SQL server intrusion and the agent deployment — planning, infrastructure setup, or something else — is not explained.
The documented point stands regardless: an AI agent can in practice run an entire intrusion chain — foothold, privilege-escalating intrusion, lateral movement, and exfiltration — quickly and without leaving AI traces on the victim machines. Defenses must therefore adapt to an attack pattern in which the techniques are old, but the tempo is new.

