AI Made a 0-Day: Noah Get the Boat

The talk opens with the view that generative AI should augment hacking rather than replace the human. The speaker frames the discussion around context, explaining how transformer models apply layered knowledge, and argues that a beginner's mind matters when prompting these tools. It separates predictable tool orchestration, where a model chooses the next scanner to run, from more open-ended generative work.

The middle of the talk walks through redacted case studies. An error in a report-generating web app became a remote code execution payload, and a model produced a working one-line version in a single attempt after days of manual research. Other cases covered a browser automation script that located form fields with validation gaps to build cross-site scripting payloads, a model suggesting an uncommon SQL technique to pull records past a web application firewall, and OCR of a captcha chained with an anti-CSRF token request to automate logins. The speaker also described gaming a search ranking system to surface weaker sites, and prompt injection used to coax a model into returning a protected file.

The closing takeaways stress that the human remains central, since the model's output is non-deterministic and results need checking. The speaker recommends keeping attack chaining creative, reserving AI for tasks that truly need it, and using private or local models to avoid content filter problems. The talk ends with a call to protect diversity of thought in how people use these tools.