Step-by-Step: How Cyber Adversaries Weaponize AI for Attack Operations

Introduction

Cyber adversaries have evolved from experimenting with AI to integrating it at an industrial scale. As of early 2026, threat groups are using generative models to discover vulnerabilities, write evasive malware, conduct autonomous attacks, and even fabricate public opinion. This how-to guide distills the latest Google Threat Intelligence Group (GTIG) findings into a step-by-step blueprint of adversarial AI tactics. Understanding these steps helps defenders anticipate and counter these emerging threats.

What You Need

• Access to premium-tier LLMs (e.g., through illicit middleware or stolen credentials)
• Automated account registration scripts to bypass usage limits
• Programming expertise (Python, C++, obfuscation techniques)
• Fuzzing and static analysis tools for vulnerability discovery
• Command-and-control (C2) infrastructure and proxy networks
• Deepfake generation tools and social media bot farms
• Software dependency analysis for supply chain attacks

Step-by-Step Process

1. Discover Zero-Day Vulnerabilities with AI

   Use generative models to perform large-scale fuzzing and code analysis. The first known case involved a criminal group that used an AI-generated zero-day exploit for a planned mass attack—only thwarted by GTIG's proactive discovery. PRC and DPRK-aligned actors have also invested heavily in AI-driven vulnerability research. Employ the model to suggest exploit paths, generate proof-of-concept code, and test against sandboxed environments.

2. Develop AI-Augmented Malware and Obfuscation

   Leverage LLMs to write polymorphic code that changes signatures on each infection. AI can accelerate the creation of obfuscation networks and decoy logic—as seen in malware linked to suspected Russian threat actors. Implement AI-generated dead code, variable renaming, and control flow flattening to evade detection. Use the model to design multi-stage loaders that blend into legitimate traffic.

3. Deploy Autonomous Malware Operations

   Integrate LLMs directly into malware to interpret system states and generate commands dynamically. PROMPTSPY is a prime example: it uses a local model to adapt its behavior, operate stealthily, and exfiltrate data without human intervention. Train a smaller, specialized model to run on the victim's device, parsing registry keys, file names, and network responses to decide next actions. This allows scaling attacks beyond manual operator capacity.

4. Use AI as Research Assistant and for Information Operations

   Employ AI to speed up every phase of the attack lifecycle—from reconnaissance to post-exploitation. In information operations (IO), generate thousands of unique social media posts, deepfake videos, and synthetic personas. The pro-Russia campaign "Operation Overload" exemplifies this: AI flooded platforms with fabricated consensus. Use prompt engineering to create varied content that avoids pattern detection.

   

5. Obfuscate LLM Access through Middleware and Account Cycling

   Threat actors now use professionalized middleware to route queries through anonymized proxy chains, and automated registration pipelines to create and cycle accounts. This bypasses usage limits and evades IP bans. Implement scripts that rotate credentials, use temporary email services, and simulate organic human interaction patterns. This infrastructure subsidizes large-scale AI misuse while reducing traceability.

6. Execute Supply Chain Attacks Targeting AI Environments
   Groups like "TeamPCP" (UNC6780) compromise software dependencies or AI model registries to gain initial access. Insert backdoors into commonly used libraries, or poison training datasets to influence outputs. Once inside an AI pipeline, adversaries can steal proprietary models, corrupt inference results, or pivot to victim networks. Use typosquatting and dependency confusion attacks against machine learning packages.

Tips for Defenders

• Deploy AI-based detection: Use ML models to spot AI-generated code patterns and obfuscation.
• Monitor anomalous LLM API usage: Look for high-volume, automated queries from diverse IPs.
• Harden supply chains: Verify checksums of AI libraries and monitor for dependency changes.
• Behavioral analysis: Focus on autonomous malware's underlying logic rather than static signatures.
• Share intelligence: Join threat-sharing communities to stay ahead of new AI-enabled techniques.