Hacking with Skynet: How AI is Empowering Adversaries

Gavin Klondike opens by framing AI as a double-edged tool: the machine learning that helps defenders with log analysis and anomaly detection can also widen what attackers can do. He separates artificial intelligence, machine learning and deep learning, then splits the talk into two halves, offensive AI tooling and attacks on machine learning systems.

The first half covers how adversaries could apply AI to social engineering, using generated faces and synthetic voices for fake personas, and to sandbox evasion, where malware checks host features such as process lists and user counts against a threshold. He also covers automated network exploitation frameworks that chain scanners and Metasploit, and the Death Star project that automates a path to domain admin. The second half covers adversarial machine learning: robustness failures, evasion of a naive Bayes spam filter and a commercial antivirus, poisoning of online learners, and data leakage from overfit models. Each attack is paired with possible defenses such as adversarial training, defensive distillation and longer retraining windows.

The closing takeaway is that adversarial examples often fail to transfer between models, so stickers and similar tricks are less practical than headlines suggest, while automation and AI-assisted tradecraft are likely to speed up attackers. Practitioners should treat model inputs, training data and retraining pipelines as attack surface, and watch for automated tooling in intrusions.