On September 10, 2026, Anthropic disclosed one of the most illustrative examples of generative AI use in a real cyber operation. The group GTG-20006, which the company assesses as linked to Russian state cyber espionage, used Claude not just for writing code or consultations. AI was integrated into nearly the entire attack cycle: reconnaissance, phishing preparation, infrastructure management, handling stolen data, and maintaining access to already compromised systems.
The main takeaway for business is much broader than this specific attack: AI is beginning to reduce the cost, time, and number of specialists needed for a complex cyber operation. While defenders could previously buy time by detecting a specific version of malware, attackers can now automate its modification and redeployment.
What Anthropic found
In the September Threat Intelligence Report Anthropic describes a group under the internal designation GTG-20006. The company says its own attribution aligns with public reporting on the group Midnight Blizzard.
Midnight Blizzard is also known as APT29, Cozy Bear, and The Dukes. The UK's National Cyber Security Centre and partner intelligence services assess APT29 as a cyber espionage group almost certainly linked to Russia's Foreign Intelligence Service.
According to Anthropic, GTG-20006 conducted operations against government, military, intelligence, and diplomatic organizations, think tanks, and defense-sector enterprises. The investigation materials mention more than 20 organizations, primarily in Ukraine and Europe.
How Claude was used in cyberattacks
The most important aspect of this case is the level of AI integration. Claude did not perform one isolated task. Anthropic describes a system in which AI was involved at virtually every stage of the operation.
| Stage | AI role |
|---|---|
| Reconnaissance | Analysis of public data, identification of email and remote services in use, preparation of target lists. |
| Initial access | Assistance in creating and operating infrastructure for phishing campaigns and other attacks. |
| Infrastructure | Researching and registering domains, configuring servers and mailing workflows. |
| Operating within the system | Performing individual actions in the compromised environment under operator control. |
| Data theft | Extracting, organizing, and processing hundreds of gigabytes of information. |
| Maintaining access | Automating individual mechanisms that enabled continued control over compromised accounts and systems. |
| Defense evasion | Monitoring the detection of malicious components and subsequently modifying them. |
In Anthropic's wording, AI in such operations is gradually moving from the role of assistant — a human helper — to the role of orchestrator, that is, a system capable of coordinating many related actions.
This aligns with the broader trend in the development of agent harness — an AI agent execution environment, where the model receives tools, memory, access to external systems, and the ability to run sequences of actions. In a legitimate business process, this makes it possible to automate routine operations. In the hands of an attacker, the same architectural idea can be used to automate cyberattacks.
AI started automatically modifying malware
One of the most important details of the report concerns the interaction between the attack infrastructure and defensive systems.
According to Anthropic, AI agents monitored whether security tools detected the malware used by the group. If a new version started being detected, the system could initiate its modification and rebuilding. The cycle continued until a variant was obtained that existing defenses did not detect.
This changes the traditional economics of cyber defense.
Previously, detecting a new malicious tool increased the cost of an attack for the attacker. AI makes it possible to partially automate the “detected → modify → test → redeploy” cycle, thereby reducing this advantage for the defender.
This does not mean that antivirus software and signature-based detection have stopped working. But static defenses alone are becoming insufficient. Behavior analysis, identity management, activity logging, and anomaly detection are becoming increasingly important.
Why the report pays so much attention to Ukraine and drones
One of GTG-20006's main targets was Ukrainian government bodies and technologies related to military drones.
Anthropic reports that the group scanned email services and remote access systems of more than two dozen Ukrainian government organizations.
In addition, the mailboxes of at least two drone component manufacturers were mass-exfiltrated. In one incident, the attackers obtained a private SDK for a drone computer vision system and then spent several days analyzing the stolen materials.
They examined the product architecture, equipment used, components, supplier dependencies, and information about a product that had not yet been announced.
This is an illustrative example of why an attacker needs AI after a successful breach. The problem is not only to obtain data, but also to quickly make sense of hundreds of gigabytes of files, emails, source code, and technical documentation. An LLM greatly simplifies such analysis.
Attacks via hotel Wi-Fi and WhatsApp
The investigation also shows another important principle of modern attacks: attackers can target not the main organization, but a weaker link around it.
Anthropic found the compromise of at least three hotel Wi-Fi providers. After gaining administrative access, the attackers changed DNS settings and used hotel infrastructure to further identify and attack guests of interest.
Special attention was paid to people connected to Ukraine, including officials and representatives of drone manufacturers.
Microsoft separately investigated this method, naming it CaptiveCrunch.
In another scenario, attackers took over WhatsApp accounts and linked controlled devices as additional ones. Anthropic reports at least two former senior Ukrainian officials who became targets of such an operation.
Hundreds of thousands of records and Microsoft 365 data
GTG-20006 did not operate only against Ukrainian targets.
In one African country, the group, according to Anthropic, gained access to a government technology agency. After compromising the VPN and central account server, the following were stolen more than 300,000 records from the national identification system and commercial registry data on more than 500,000 companies.
Separate infrastructure was used for attacks on Microsoft 365. The Embassy Kit framework developed by the group was used to steal tokens and gain access to the cloud email of diplomats and government employees.
Anthropic reports access to email data from at least eight organizations.
Why this case matters for ordinary businesses
It may seem that such threats concern only governments, defense companies, and intelligence agencies. This is a dangerous oversimplification.
The technical shift is not in the choice of specific victims, but in reducing the cost of a sophisticated attack.
AI can speed up work that previously required dedicated specialists:
- collect and classify information about a company;
- tailor phishing to a specific employee;
- analyze unfamiliar infrastructure;
- review documentation and source code;
- process large volumes of stolen information;
- automate repetitive actions;
- coordinate several parallel tasks.
In other cases from Anthropic's September report, it already describes multi-agent systems that conducted reconnaissance, exploitation, and information theft in parallel over an extended period.
For a company, this means that an attacker's ability to scale becomes less dependent on the size of their team.
The same problem exists within corporate AI agents
There is an important parallel between this investigation and the usual adoption of AI in business.
When companies connect an AI agent to email, CRM, Google Drive, a customer database, GitHub, or internal infrastructure, they effectively create a system that can also read data and perform actions.
Therefore, security cannot be solved with a single system prompt like “do not do anything dangerous”.
In CenterAI's article on AI agent security a different approach is used: separate accounts, least-privilege permissions, human approval of critical operations, an activity log, and the ability to stop the agent immediately.
For autonomous systems, the boundaries of the execution environment are especially important. The analysis of incidents involving autonomous AI agents already shows why the combination of the internet, tools, and insufficient restrictions creates an entirely different level of risk than a regular chat with an LLM.
What needs to change in company security
AI does not replace traditional cybersecurity, but it makes weak points much cheaper to exploit. Therefore, the main task is not to guess whether a particular attacker uses AI, but to reduce the number of options available to them.
- Enable multi-factor authentication. Especially for email, VPN, cloud services, GitHub, and administrative interfaces.
- Reduce user and service account privileges. A single stolen account should not grant access to the entire infrastructure.
- Monitor tokens and API keys. Secrets should not be stored in public code, documents, client applications, or shared folders.
- Track new devices and sessions. Unusual device registration or a login from a new environment should trigger a review.
- Protect more than just the central system. Suppliers, contractors, SaaS services, and other supply chain elements may become an easier entry point.
- Log the actions of your own AI agents. If an agent has access to CRM, email, or customer data, the company should understand exactly what it has read, changed, or sent.
- Keep humans involved in critical decisions. Sending payments, deleting data, changing permissions, publishing, and transferring confidential information should not happen without oversight.
When designing an enterprise agent, it is better to build in these restrictions already at the MVP stage. CenterAI uses the approach “process → limited MVP → testing → launch”, rather than connecting AI to the company's entire infrastructure at once.
Key takeaway from the Anthropic investigation
The GTG-20006 case demonstrates the next stage in the use of generative AI.
AI is already used for more than generating phishing emails or code. It is becoming a software layer that connects tools, data, and individual stages of a cyber operation into a single workflow.
Humans are not disappearing from the process. In the operation investigated by Anthropic, operators continued to select targets, direct the work, and analyze the results. But a significant part of repetitive technical work could be delegated to AI.
That is why the main risk is not how “smartly” a chatbot responds, but what tools, data, and permissions the model receives and how autonomously it can use them.
This principle is equally important for attackers and for companies implementing their own AI agents.
If a company plans to connect AI to corporate data, email, CRM, or external services, it makes sense to first define access boundaries and critical operations. CenterAI can conduct an AI audit of the process and design a limited MVP with control over agent access and actions.
