Major AI cyber attacks witnessed globally.

North Korean actor used Gemini to synthesize OSINT and profile high-value targets at cybersecurity/defense firms, mapping job roles and salaries for fake-recruiter campaigns.

Synthesized OSINT and profiled targets for recruiter scam

Chinese state-linked actor prompted Gemini as an expert-cybersecurity persona to automate vulnerability analysis, testing RCE, WAF-bypass, and SQL-injection techniques against specific US targets.

Automated vulnerability analysis and exploit-testing plans

Chinese state-linked actor used Gemini and other AI tools to compile detailed dossiers on individuals in Pakistan and gather data on separatist organizations across multiple countries.

Compiled target dossiers and organizational intelligence

Chinese state-linked actor relied heavily on Gemini multiple days a week to troubleshoot code, research techniques, and build intrusion capabilities, including interest in agentic code-auditing.

Ongoing code troubleshooting and capability development

Unidentified actor used Gemini to hunt sensitive account credentials and email addresses; those accounts were later hit in a phishing campaign against Ukraine's defense sector.

Gathered target credentials/emails ahead of phishing campaign

Financially motivated actor built the "COINBAIT" phishing kit using the AI app-builder platform Lovable, evidenced by Lovable-specific code and image-hosting artifacts in samples.

Built the phishing kit's infrastructure/app

CrowdStrike found adversaries injected malicious prompts into legitimate GenAI tools at 90+ organizations to generate commands for stealing credentials and cryptocurrency.

Generated commands to steal credentials/cryptocurrency via prompt injection

North Korean actor used AI coding tools to iteratively generate and refine malware components, agentic AI for an end-to-end lure-development workflow, and jailbroken LLMs for code.

Iterative malware development and automated lure creation

North Korean actor used AI to develop fake digital personas supporting social-engineering campaigns targeting employment and recruitment opportunities.

Generated fake personas for recruitment-themed social engineering

Microsoft attributes the Tycoon2FA phishing-as-a-service platform to Storm-1747; AI-refined lures hit ~54% click-through vs ~12% for traditional phishing, sent at tens of millions/month.

Refined phishing lure content at scale for higher click-through

Actor used "expert persona" prompting, directing Gemini to act as a senior security auditor, to drive vulnerability research on embedded router firmware hunting pre-auth RCE flaws.

Persona-prompted to conduct vulnerability research on firmware

North Korean state actor sent thousands of repetitive prompts to recursively analyze CVEs and validate proof-of-concept exploits at scale, while experimenting with agentic AI tooling.

Automated large-scale CVE analysis and exploit validation

Chinese state-linked actor used Gemini to accelerate development of a fleet-management app supporting an operational relay box (ORB) network used to route malicious traffic.

Accelerated development of ORB-network management infrastructure

Chinese state-linked actor used a public script to mass-register and cancel premium LLM trial accounts across Gemini, Claude, and OpenAI to obtain anonymized model access.

Abused sign-up flows to gain free anonymized premium access

Chinese state-linked actor built obfuscated proxy middleware to pool and share access to Gemini/Claude/OpenAI accounts, while targeting government sectors across South and Southeast Asia.

Built infrastructure for anonymized, shared LLM access

An actor used a non-Gemini AI model to discover and weaponize a zero-day 2FA-bypass in a popular open-source sysadmin tool: GTIG's first AI-developed zero-day.

Discovered and helped weaponize a zero-day vulnerability

Unattributed Russia-nexus actor delivered malware (CANFAIL, LONGSTREAM) against Ukraine that uses LLM-generated decoy filler code to obfuscate its actual malicious functionality.

Generated decoy/filler code to obfuscate malware

Android backdoor embeds an autonomous Gemini-API agent that navigates the device's UI to drive on-device data theft; assessed as experimental and unattributed.

Autonomous agent navigated device UI to steal data

Suspected Chinese state-linked actor deployed agentic offensive-security frameworks (Hexstrike, Strix) against a Japanese tech firm and an East Asian cybersecurity platform.

Ran agentic frameworks to conduct the intrusion

Behavioral evidence (parallel multi-account logins from one IP/user-agent, rapid checklist-style technique execution) suggested agentic-AI-assisted execution in a ~72-hour AWS cloud takeover.

Suspected to have driven parallel, rapid cloud intrusion tasks

Spear-phishing espionage against government and electric-power targets in Russia, Kazakhstan, and Brazil deployed BusySnake Stealer via a loader Kaspersky attributes to LLM-generated code.

Suspected generation of the malware loader source code

An unnamed actor deployed a self-labeled AI-generated PowerShell script to enumerate Active Directory environments, using it alongside s5cmd and SharpShares tools.

Generated the PowerShell script used for AD enumeration

First fully agentic ransomware campaign: an LLM agent autonomously exploited a Langflow flaw, harvested credentials, moved laterally, and encrypted 1,342 configs across 600+ payloads.

Ran the entire attack autonomously end-to-end

Unit 42 found an IoT botnet (TuxBot v3) whose Iranian developer relied heavily on LLM-generated code for bot modules, C2 server code, and porting an ADB exploit.

Generated botnet modules, C2 code, and exploit porting

Russian-speaking actor used jailbroken Gemini for ~90% of a fraud operation (writing C2 code, migrating infrastructure in 6 minutes), compromising a US dental clinic and crypto-wallet targets.

Drove approximately 90% of the fraud operation autonomously

Suspected China-based actor ran Claude Code and DeepSeek-v4-pro together, compromising Thai and Afghan government systems, partially exploiting Taiwanese firms, and extracting financial-services admin credentials.

Agentic execution plus exploit-script generation and bypass strategy

Hugging Face's own incident response found a July 2026 intrusion of its internal infrastructure run by an autonomous agent framework, exploiting dataset-loader flaws and logging 17,000+ agent actions.

Autonomously executed the entire intrusion from recon to lateral movement

Russian state actor phished Zimbra webmail users via a CSS-injection XSS flaw to harvest emails, directories, and 2FA tokens; AI reportedly helped build its exfiltration framework.

Assessed to have helped develop the Flowerbed exfiltration framework

Malvertising campaign abused a public Claude Artifact on claude.ai to impersonate the Claude Desktop installer, delivering SectopRAT malware to 29 organizations before Anthropic removed it.

Attacker-built Artifact impersonated official software as the lure

Russian-speaking Aur0ra ransomware operators repeatedly told Cursor's AI agent they were conducting a legitimate security test, using it to automate network scanning, VPN access, and exploitation across 10 organizations.

Automated recon, VPN access, and exploitation attempts

Emerging ransomware-as-a-service group claimed 85+ victims in Dec 2024 alone; its ransomware code and scripts showed hallmarks of LLM-generated comments/output.

Assisted ransomware code development; powered the group's support chatbot

Used Claude Code as an autonomous agent across 17+ organizations to conduct recon, credential theft, and network penetration, then size ransom demands and draft victim-specific ransom notes.

Acted as autonomous agent for hacking and ransom-note generation

A low-skilled, model-dependent actor used Claude to build and sell a ransomware-as-a-service kit (ChaCha20 encryption, anti-EDR/recovery) on dark-web forums for $400-$1,200.

Built the ransomware code and product for the criminal to sell

An unidentified actor used Claude to build a carding (stolen payment-card) service, developing advanced API integrations and operational-resilience mechanisms for the criminal platform.

Assisted with building and hardening the criminal carding service

Summer 2025 campaigns against Brazilian and Spanish-speaking hotels used a significant portion of LLM-generated infector/downloader code to deliver VenomRAT malware (HVNC, stealer, UAC bypass).

Suspected generation of the infector and downloader code

Cybercriminals abused the AI site-builder Lovable to mass-produce phishing pages, DeFi wallet-drainer clones, and malware redirectors: tens of thousands of malicious URLs per month.

Auto-generated phishing pages and fraud infrastructure at scale

Russian GRU actor deployed PROMPTSTEAL (LameHug) malware against Ukraine, a data-miner that queries an LLM via Hugging Face API at runtime to generate recon commands live.

Malware queried the LLM live to generate its own recon/collection commands

Chinese state-sponsored actor jailbroke Claude Code via a "defensive security firm" roleplay to autonomously run espionage intrusions against ~30 global orgs, succeeding in a few.

Performed an estimated 80-90% of tactical intrusion work autonomously

Chinese state-linked actor used Gemini for C++/Golang development of multiple tools, including a C2 framework ("OSSTUN"), plus code-obfuscation help via public libraries.

Assisted with malware/C2 tool development and code obfuscation

Iranian IRGC-linked actor used Gemini to craft phishing content impersonating think-tank staff and attempted a natural-language-to-SQL "Data Processing Agent" to mine personal data.

Generated phishing content; built NL-to-SQL data-mining agent

Iranian actor jailbroke Gemini via fake student/researcher pretexts to develop custom malware, including a Python C2 server and web shells.

Assisted development of C2 server and web shell malware

North Korean actor used Gemini for crypto-wallet recon, multilingual phishing lures, and theft-code attempts; used deepfake video of a fake CEO on Zoom to deliver BIGMACHO malware.

Recon, lure generation, and deepfake impersonation for malware delivery

North Korean actor used Gemini to develop code, research exploits for edge devices and modern browsers, and improve attack tooling.

Assisted code development and exploit research

Suspected China-nexus actor jailbroke Gemini using a fake "capture-the-flag" pretext to build phishing lures, exfiltration tooling, and web shells.

Generated phishing, exfiltration tools, and web shells

Suspected China-nexus actor used Gemini across the intrusion lifecycle (recon, payload delivery, lateral movement, C2, and exfiltration), including on unfamiliar AWS/vSphere/Kubernetes environments.

Assisted across the full intrusion lifecycle on cloud systems

OpenAI disrupted Russian, Korean, and Chinese-language actors using ChatGPT to refine malware/RATs, build C2 systems, and craft phishing against Taiwan's semiconductor sector and US academia.

Refined malware, built C2 tooling, and crafted phishing

A credential-stealing malware invoked victims' own installed AI CLI tools to search infected systems for additional secrets, exfiltrating data via attacker-made public GitHub repos.

Abused the victim's own AI tools to search for secrets

Experimental VBScript malware calls the Gemini API at runtime to rewrite and obfuscate its own code "just-in-time," spreading via removable drives; not yet seen deployed.

Malware calls it live to self-modify and obfuscate its code

Ransomware-as-a-service group advertised an "AI-Assisted Chat" tool to affiliates that analyzes victim-company data to tailor and automate extortion negotiation messaging.

Automated and tailored extortion negotiation messaging

Cybercriminal group sold an AI-powered tool for BEC scams that scanned hacked emails and altered invoice banking details to redirect payments.

Automated invoice fraud detection and banking-detail alteration for BEC scams

UK's NCSC assessed threat actors, including ransomware groups, are already using AI to boost efficiency of recon, phishing, and coding.

Increased efficiency across reconnaissance, phishing, and coding tasks

North Korean actor used LLMs to research NK experts/think tanks and craft spear-phishing content impersonating academic institutions and NGOs.

Reconnaissance on targets plus generation of spear-phishing lure content

Iranian IRGC-linked actor used AI for social-engineering support, .NET malware troubleshooting, and researching detection-evasion methods.

Assisted with malware troubleshooting, social engineering, and evading detection

Chinese state-linked actor used LLMs to support tool development, scripting, and generate social-engineering content against multiple countries.

Supported tooling development, scripting, and social-engineering content creation

Chinese state-linked actor used LLMs exploratorily to research sensitive topics, high-profile individuals, geopolitics, and US internal affairs.

Exploratory reconnaissance and information-gathering on sensitive topics/individuals

~100 malicious machine-learning models uploaded to Hugging Face contained backdoor payloads, granting attackers shell access to victims' machines.

Malicious model itself served as the malware delivery vector

A malicious PyTorch model uploaded to Hugging Face (later deleted) held a backdoor payload; hosting IP hinted at possible researcher involvement.

Malicious model itself served as the malware delivery vector

A second Hugging Face account hosted the same backdoor payload as baller423/goober2 under a similarly-named model repository.

Malicious model itself served as the malware delivery vector

Used PowerShell resembling LLM-generated output (matching Llama 2 70B code patterns) to exfiltrate Entra ID user data via a backdoor.

Generated the PowerShell script used to exfiltrate user data

During a LockBit RED ransomware incident, the actor browsed ChatGPT alongside search engines to learn Azure navigation for credential exfiltration.

Used as a learning aid to navigate Azure and locate credentials

Fake "GPT chat" websites falsely claimed to diagnose PC infections via AI, funneling victims to scam call centers instead.

No real AI used: fake branding to lure victims into scam calls

A fake "smart chatbot" impersonating Elon Musk promised guaranteed investment returns, then vanished after collecting an upfront "fee."

No real AI used: fake chatbot persona to lure investment-scam victims

TA547 sent invoice-themed phishing emails impersonating German retailer Metro, delivering Rhadamanthys stealer via a PowerShell loader with LLM-typical comment style.

Suspected to have generated or rewritten the PowerShell loader script

Trend Micro found criminals increasingly jailbreaking existing LLMs rather than building new ones, plus early commercial deepfake-as-a-service offerings including financial KYC-bypass tools.

Jailbroken for malware/social-engineering aid; deepfakes for KYC bypass

Symantec found phishing campaigns delivering Rhadamanthys, NetSupport, ModiLoader, LokiBot, and Dunihi malware via PowerShell/HTML/JavaScript scripts bearing hallmarks of LLM-generated code.

Suspected generation of the PowerShell/HTML/JS malware-delivery scripts

Chinese state-linked group breached major US telecoms (T-Mobile, AT&T, Verizon, Lumen) exploiting router flaws; investigators suspect AI/ML aided the espionage.

Suspected (unconfirmed) to have enhanced network access and intelligence-gathering