AnswersPublished22 sources
How AI Is Changing Cyberattacks: From AI-Generated Malware to LLM-Driven Intrusions
AI is making existing cyberattacks faster, cheaper and more adaptable rather than replacing conventional methods. FunkSec and the 2025 RevengeHotels campaign show AI assisted malware development, while LameHug demonstrates a more consequential shift: malware querying an LLM at runtime to generate reconnaissance and...
Published byEdited with GPT-5.6 LunaImages generated with GPT Image 1.5
Research answer
Key takeaways
- AI is making existing cyberattacks faster, cheaper and more adaptable rather than replacing conventional methods.
- FunkSec and the 2025 RevengeHotels campaign show AI assisted malware development, while LameHug demonstrates a more consequential shift: malware querying an LLM at runtime to generate reconnaissance and data collectio...
- The practical defensive lesson is to look beyond static signatures for unusual scripting, outbound LLM/API traffic, credential theft and data staging.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: How is artificial intelligence becoming deeply embedded in cyberattacks, according to Kaspersky and other cybersecurity researchers, and wha. Article summary: AI is becoming an operational layer across the attack chain: it helps criminals write and revise code, tailor phishing, automate reconnaissance, and vary payloads at machine speed. The clearest current evidence supports . Topic tags: general, general web, user generated. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fa
Artificial intelligence is becoming an operational layer in cyberattacks. Criminal and state-linked operators can use large language models to draft malware components, produce phishing content, troubleshoot scripts and generate new variations faster than traditional workflows allow. But the clearest evidence so far supports a more precise conclusion: AI is accelerating and adapting familiar attack methods, not making every attack autonomous.
The main change: AI compresses the attack cycle
Most intrusions still depend on recognizable techniques such as phishing, malicious scripts, remote-access trojans and data theft. AI changes the economics of those techniques by reducing the time and expertise needed to produce them.
An attacker can use an LLM to create an initial script, translate or personalize a lure, adjust code after testing, and generate multiple variants for different targets. That makes experimentation cheaper and can shorten the time between a campaign’s initial idea and its deployment. The result is a higher potential volume of credible attacks and a smaller response window for defenders.
The distinction matters. Calling every attack that uses an AI-themed lure an “AI attack” overstates the evidence. The strongest examples show several different levels of AI involvement, from code assistance to live command generation.
FunkSec shows AI-assisted malware development
Kaspersky’s research on FunkSec provides one example of AI-assisted malware construction. The Rust-based ransomware combines encryption, local data exfiltration and self-cleanup in a single executable. Kaspersky also reported signs of large-language-model code synthesis in related capabilities, including a password generator and a basic DDoS tool. 28
39
This does not mean that an LLM independently designed, deployed and operated the entire campaign. It shows that generative AI can contribute to functional malware components, helping attackers assemble a more consolidated toolkit with less manual development.
RevengeHotels: AI improves a conventional phishing chain
The 2025 RevengeHotels campaign illustrates how AI can strengthen an established intrusion method rather than replace it. Kaspersky observed attacks against hotels in Brazil and Spanish-speaking markets that used phishing emails disguised as reservation requests, invoices or job applications. Victims were directed to sites mimicking document-storage services, where scripts ultimately delivered remote-access trojans such as VenomRAT. 30
31
36
Kaspersky said a significant portion of the initial infector and downloader code appeared to have been generated by large-language-model agents. 28 The campaign still relied on familiar social engineering, script execution and RAT delivery. AI’s contribution was to help produce and vary parts of the infection chain, potentially allowing the operators to iterate across regions and targets more efficiently.
That is why “AI-powered” can be a misleading label when used without explanation. In this case, AI-assisted code generation was embedded inside a conventional phishing operation.
LameHug brings an LLM into the attack itself
The most important technical development in the available evidence is LameHug, also referred to in some reporting as PROMPTSTEAL. Unlike malware that merely appears to have been written with AI assistance, LameHug used an LLM during execution.
According to analyses based on CERT-UA reporting, the malware contacted the Hugging Face API and used Alibaba Cloud’s Qwen2.5-Coder-32B-Instruct model to generate Windows commands dynamically. Those commands could gather system information and search for Office documents for exfiltration. 2
3
7
This runtime model changes the malware’s behavior in three important ways:
- The command sequence does not have to be fully fixed in advance. The malware can request commands based on its instructions and the environment it encounters.
- The attack can use ordinary system utilities. Detection may need to focus on the behavior and sequence of commands rather than on a distinctive standalone payload.
4
- The malware creates an external dependency. A compromised system making unusual outbound requests to a language-model API can become a valuable detection signal.
LameHug is not proof that malware has become autonomous. It is evidence that an attacker can place an LLM inside the operational loop, allowing the model to help generate reconnaissance and data-collection actions in real time.
AI-themed lures are expanding the attack surface
AI is also being used as bait. From January through April 2026, Kaspersky solutions detected more than 33,300 attacks against small and medium-sized businesses in which malicious or unwanted PC software was disguised as popular AI services. Kaspersky described the figure as almost five times the comparable total from 2025. 50
51
These detections measure software masquerading as AI services. They do not establish that every payload was generated by AI or that an AI system operated each attack. The more defensible conclusion is that attackers are exploiting trust in familiar AI brands and the rapid adoption of AI tools inside businesses.
For employees, the risk is practical: a fake image generator, chatbot client or productivity assistant can become the delivery mechanism for malware. Software installation policies, verified download sources and strong endpoint controls therefore remain important even when the lure appears to be an AI product.
What the evidence says about scale and regional exposure
Kaspersky’s regional telemetry shows that web-based threats remain widespread, but those figures should not be treated as counts of AI-driven attacks. During the first half of 2026, Kaspersky said its systems blocked 5.7 million attacks from online resources in South Africa, 4.5 million in Kenya and 1.6 million in Nigeria. Türkiye recorded the highest share of users affected by web-based threats in the reported META-region dataset, at 22.8%. 14
The measures cover threats delivered through websites, email and web services. They demonstrate broad exposure to online attacks, not AI causation. Keeping that distinction clear prevents general cyberthreat statistics from being presented as evidence of autonomous or AI-generated campaigns.
Why adaptation matters more than autonomy
The immediate security risk is not necessarily a sudden transition to self-directing cyber armies. It is the accumulation of smaller efficiencies across the attack chain:
- code can be drafted and revised more quickly;
- phishing messages can be adapted for different languages, industries and pretexts;
- scripts can be debugged and repackaged faster;
- malware behavior can be varied across environments;
- reconnaissance and data collection can be adjusted at runtime.
Together, these efficiencies can lower the cost of launching an intrusion and make it easier to run more experiments. They may also help less-skilled operators imitate techniques that previously required specialist knowledge.
The risk is especially serious for organizations with limited security staff, large volumes of sensitive data or complex third-party environments. However, claims about attacks on critical infrastructure require specific evidence of compromise, impact and AI involvement. A claim by an attacker is not, by itself, confirmation of any of those facts.
Defensive priorities for organizations
The examples above point toward a layered, behavior-focused response. Organizations should continue using signatures and known-indicator blocking, but they should also monitor for activity that can reveal AI-assisted or dynamically adapting attacks.
Useful signals include:
- Python or PyInstaller processes making unexpected outbound requests to language-model services;
- base64-encoded prompts or unusual API calls from endpoints that normally do not use developer or AI services;
- command chains that combine utilities such as
systeminfo,tasklist,dsqueryor recursive file searches; - suspicious scripting through JavaScript, PowerShell or other interpreters;
- unusual credential access, document staging or outbound transfers;
- employees installing unapproved AI applications or browser tools.
Splunk’s LameHug detection guidance specifically highlights Python or PyInstaller activity that connects to the Hugging Face API and uses the Qwen model to generate commands. 4 This is a concrete example of how defenders can identify the behavior around an LLM-enabled attack, rather than waiting for a fixed malware signature.
The bottom line
AI is becoming deeply embedded in cyberattacks because it makes existing operations easier to produce, personalize and modify. FunkSec and RevengeHotels show AI-assisted malware development; LameHug shows an LLM being queried during a live intrusion; and Kaspersky’s SMB detections show that the popularity of AI itself is being used as a lure. 28
31
51
The most accurate description is therefore not “AI has replaced conventional hacking.” It is that AI is becoming an efficiency and adaptation layer across phishing, malware development, reconnaissance and data theft. Defenders should prepare for more variation and faster iteration while continuing to distinguish verified AI involvement from ordinary cybercrime wrapped in AI branding.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "How AI Is Changing Cyberattacks: From AI-Generated Malware to LLM-Driven Intrusions"?
AI is making existing cyberattacks faster, cheaper and more adaptable rather than replacing conventional methods.
What are the key points to validate first?
AI is making existing cyberattacks faster, cheaper and more adaptable rather than replacing conventional methods. FunkSec and the 2025 RevengeHotels campaign show AI assisted malware development, while LameHug demonstrates a more consequential shift: malware querying an LLM at runtime to generate reconnaissance and data collectio...
What should I do next in practice?
The practical defensive lesson is to look beyond static signatures for unusual scripting, outbound LLM/API traffic, credential theft and data staging.
Sources
me-en.kaspersky.comKaspersky uncovers the cyber threats defining the first half ...
securelist.comA new RevengeHotels campaign targets Latin America
kaspersky.comKaspersky discovers the return of RevengeHotels, leveraging AI in ...