AI systems introduce behavioral risks that traditional security practices like secure coding and penetration testing do not fully address.
AI's probabilistic nature creates unique risks like bias and toxicity, while new attack vectors like prompt injection expand the traditional attack surface.
Tools like SAST and DAST, designed for explicit code, cannot adequately assess emergent AI behaviors like bias, hallucinations, or malicious prompt following.
AI models can be used to find and explain security issues in code, accelerating the race between attackers discovering exploits and defenders performing remediation.
The software development lifecycle must expand to a trustworthy AI SDLC, integrating concerns like data quality, model fairness, and continuous monitoring from the start.
A comprehensive GenAI security model must address four distinct areas: data governance, model validation, secure source code, and inference-time behavior monitoring.
Trustworthy AI is built on three pillars: responsible AI for fairness and transparency, AI security for robustness, and AI privacy for data protection.
OWASP provides practical frameworks like the Top 10 for LLMs and the AI Testing Guide to help organizations translate abstract principles into concrete practices.
The OWASP AI Testing Guide provides a structured, risk-driven strategy for testing each layer of an AI system, from data and infrastructure to the model and application.
The OWASP AI Maturity Assessment helps organizations evaluate their readiness for AI by assessing governance, policies, skills, and operational controls through a structured journey.
Building trustworthy AI is not just a technical challenge but an organizational one, requiring new skills, processes, and a commitment to continuous monitoring.
Jobs that call for the skills explored in this talk.