Geoff McDonald
Geoff McDonald is a data scientist, security researcher, and leader at Microsoft Defender for Endpoint. He leads machine learning and data science specializing in the latest AI trends and building machine-learning and automation solutions to prevent, detect, and disrupt cyber-attacks. He has a passion for machine learning, reverse-engineering, ai safety, programming tools for reverse-engineering and vulnerability fuzzing, and foosball. You can find some of his tools and hobby projects on his personal website https://www.split-code.com or GitHub at https://github.com/glmcdona.
2026 Talk
Talk Title: Fully autonomous AI attacks, and worms that vibe-code themselves - update on where we are and what to expect
Talk Abstract:
A critical threshold has been crossed in cybersecurity. In July, we observed the first known AI-orchestrated complex ransomware attack (JADEPUFFER). This was immediately followed by two alarming industry disclosures: an OpenAI model escaping its training sandbox to autonomously hack Hugging Face for test answer keys, and Anthropic revealing that its Claude models inadvertently breached three real-world organizations after reaching the open internet during misconfigured cybersecurity evaluations.
This talk provides a stark update on where GenAI currently stands, and where it is rapidly heading, against critical cybersecurity thresholds. With frontier capabilities doubling every 127 days, and capable open-weight models lagging in attack orchestration capabilities by mere months, cyberattacks are now scaling based on compute budgets rather than human skill. First, we will dissect the mechanics of autonomous AI attack orchestration, what lab cyber range benchmarks show, and review these notable recent incidents. Then, we will discuss the imminent threat of vibe-coded, self-improving worms that turn every infected node into an independent malware development environment, sparking a rapid cycle of natural selection evolving worms.
The new math for defenders is uncompromising:
AI attack orchestration allowing scale limited primarily by budget
➕ Open-weights models making it trivial to remove post-training safety layers
➕ METR demonstrating GenAI is on an exponential curve, doubling every 127 days
🟰 Unprecedented enterprise risk.