Security in the Age of Agentic AI | Vasu Jakkal, Microsoft Security | RSAC 2025

### Overview In her keynote at RSAC 2025, Vasu Jakkal, Corporate Vice President of Microsoft Security, examines the transformative rise of agentic AI and its profound impact on cybersecurity. Jakkal highlights that as interactive, autonomous AI agents become ubiquitous across organizations and daily life, the security landscape faces an unprecedented escalation in threat volume and velocity. To meet this moment, defenders must address two deeply interconnected imperatives: leveraging AI agents to defend networks at machine speed and scale, and securing AI systems themselves from the ground up.

### Key Topics - **Escalating Threat Landscape**: Attack metrics have intensified drastically, rising from 4,000 to 7,000 password attacks per second (600 million daily). Attackers can gain access to full data within 72 minutes or less after a phishing link click, while tracked adversary groups grew from 300 to 1,500 over the past year. Furthermore, insider risk remains significant, with 20% of data breaches involving insiders. - **Autonomous Evolution of AI**: Over the next two years, AI systems are anticipated to advance from Level 0 autonomy (performing explicit tasks and handling alert drudgery) to Level 3 autonomy (generating sub-goals, autonomously altering models, and taking independent actions). - **Two Interconnected Security Mandates**: Security leaders must pursue a dual strategy: applying AI for security defenses while implementing rigorous security for AI architectures. - **Human Role and Governance**: Human agency remains central as workforce roles transition from manual task execution to directing, guiding, and governing autonomous agents.

### Technical Details - **Using AI for Security**: Microsoft trains security models on 84 trillion daily signals, tuning them for specific security skills to prevent hallucinations. Current applications include threat investigations, incident reporting, reverse engineering, vulnerability patching, triage of billions of phishing alerts, and dynamic conditional access enforcement. Future capabilities envision moving from responsive to predictive defense, identifying data risk at point-of-creation, and achieving cross-domain superhuman competency. - **Security for AI**: As adversaries leverage AI for malware variants, social engineering, vulnerability discovery, intelligent password cracking, and deepfakes, AI architectures must enforce security by design and default. Essential controls include giving agents explicit identities governed by Zero Trust and least privilege, preventing data leakage and oversharing, and mitigating specific attacks such as prompt injections, LLM model poisoning, model abuse, and wallet abuse. - **Shift in Governance**: Because agents operate dynamically, governance must evolve from lab-based static verification to dynamic, probabilistic verification across the entire lifecycle (onboarding, offboarding, and decommissioning), including embedding security sub-agents within primary agents.

### Takeaways - **Trust and Responsible AI**: Foundational AI trust requires adherence to core pillars: fairness, inclusion, transparency, accountability, privacy, security, and safety. - **Emphasizing Human Strengths**: The shift to agentic workflows frees defenders to focus on innovation, creativity, growth mindset, and empathy. - **Diversity and Collaboration**: Confronting diverse global adversaries requires cognitive diversity and cross-industry collaboration, reinforcing the principle that security is fundamentally a team sport.