Machine-generated domains and query traffic¶

Attack pattern¶

Domain generation algorithms have always had a tell. They produce names no person would choose, and classifiers learned to read that lexical signature years ago. A language model does not have the tell, because it produces names that look like names somebody registered on purpose. The same shift applies on the query side, where the goal is to match a benign distribution rather than to hide a packet: a tunnel shaped to resemble ordinary resolver traffic passes a volume threshold without concealing anything at all. Reputation systems that retrain on their own feedback are the third target, and they fail in the way any retrained classifier fails, which is covered in its own right under model poisoning. None of these techniques is new. What moved is the cost of producing ten thousand plausible instances of one.

1. AI/ML-Augmented Attacks [OR]

    1.1 Evasion Techniques [OR]

        1.1.1 Reputation System Poisoning [OR]
            1.1.1.1 DNS query pattern manipulation
            1.1.1.2 Behavioural model contamination
            1.1.1.3 Feedback loop exploitation

        1.1.2 Query Obfuscation [OR]
            1.1.2.1 GAN-generated benign-looking queries
            1.1.2.2 CDN traffic mimicry
            1.1.2.3 Legitimate domain spoofing

        1.1.3 Adaptive Attacks [OR]
            1.1.3.1 Reinforcement learning for evasion
            1.1.3.2 Genetic algorithm optimisation
            1.1.3.3 Transfer learning across networks

    1.2 Phishing Automation [AND]

        1.2.1 Domain Generation [OR]
            1.2.1.1 LLM-generated homograph domains
            1.2.1.2 Context-aware typosquatting
            1.2.1.3 Cultural adaptation algorithms

        1.2.2 Infrastructure Management [OR]
            1.2.2.1 Dynamic DNS fast-flux networks
            1.2.2.2 Automated certificate procurement
            1.2.2.3 Multi-CDN abuse for resilience

        1.2.3 Target Identification [OR]
            1.2.3.1 NLP-based brand monitoring
            1.2.3.2 Social media sentiment analysis
            1.2.3.3 Employee behaviour prediction

Cheap plausibility¶

• Classical detection of generated domains reads lexical features, and a name written by a language model reads like one a person would have registered.

• Reputation systems learn from their own feedback, which makes them the same target as any other retrained classifier, and they fail the same way.

• Query obfuscation aims at the distribution rather than the packet. Traffic shaped to match the benign shape defeats a volume threshold while hiding nothing.

• Automated certificate issuance supplies the padlock, so a lure registered an hour ago carries the same visual assurance as anything else.

• The economics moved rather than the techniques. Homographs, typosquats and fast-flux all long predate the tooling; producing them at volume is the change.

• Much of this branch is demonstrated rather than widely observed, and the risk table prices it at medium for that reason.