When AI Agents Go Rogue

AI systems are becoming increasingly capable of planning, executing code, using tools, accessing sensitive information, and coordinating with other agents. As these capabilities grow, so does the potential for AI to be misused or to act beyond its intended boundaries.
Join John Sotiropoulos for an interactive red and blue team workshop exploring the changing face of adversarial AI inspired by the reported Mexican government breaches and this summer's agentic sandbox escapes.
You’ll work through practical scenarios from both the attacker and defender perspectives, making decisions, analysing evidence, and testing how different security controls can change the outcome.

Meet Your Instructor

John Sotiropoulos is the author of Adversarial AI Attacks, Mitigations, and Defense Strategies (Packt) and founder of Deep Cyber, a UK AI security consultancy serving regulated financial services and critical national infrastructure clients. He is a Board Director of the OWASP GenAI Security Project, chairs the OWASP Top 10 for Agentic Applications and co-leads the OWASP Agentic Security Initiative.

What this workshop is really about

This workshop goes beyond theory and explores what happens when AI becomes an active participant in an attack or when an AI agent starts operating beyond the boundaries it was given.

You’ll explore two key scenarios:

• AI as the attacker’s operational partner: Examine how AI can help accelerate an intrusion, make decisions, use compromised access, and turn an initial foothold into an ongoing capability.
• Rogue AI agents: Explore how agents can pursue objectives beyond their authorised remit and how permissions, shared information, memory, tools, and coordination can increase the impact.

You’ll take both red team and blue team perspectives, identify weaknesses, investigate suspicious activity, and determine where defenders should intervene.

What you'll be able to do by the end of workshop

By the end of the workshop, you’ll have practical experience in:

• Analysing AI-assisted attack scenarios.
• Red teaming AI agents and agentic systems.
• Identifying where permissions, credentials, tools, and shared information can create security risks.
• Investigating suspicious agent behaviour.
• Evaluating monitoring, approval, access-control, and containment strategies.
• Making defensive decisions when an AI system begins operating outside its authorised boundaries.

Who should attend

This workshop is designed for:

• Security practitioners
• Red teamers and blue teamers
• AI/ML engineers
• Security engineers and architects
• AI security professionals
• Technical leaders responsible for deploying AI systems
• Cybersecurity professionals interested in emerging AI threats

What you’ll learn

• How AI is changing the nature of adversarial activity.
• The difference between AI-assisted attacks, manipulated instructions, and rogue agent behaviour.
• How tools, credentials, memory, and agent coordination can expand an AI system’s attack surface.
• How to identify and investigate suspicious agent activity.
• How to evaluate permissions, approvals, monitoring, and containment controls.
• How lessons from reported AI security incidents can be applied to systems you build, assess, or defend.
• How to approach red teaming for agentic and multi-agent systems.
  
  Prerequisites (Knowledge)
  
  - Basic understanding of cybersecurity concepts.
  - Familiarity with AI systems or AI agents is helpful but not required.
  - No previous red or blue team experience is required.
  
  Prerequisites (Software/Hardware)
  - A laptop or desktop computer with a reliable internet connection.
  - A modern web browser.
  - No specialized security software or hardware is required.
  - All practical exercises will take place in simulated environments.
  
  We'll Provide
  - Access to the live online workshop.
  - Hands-on simulated environments for the practical exercises.
  - Guided red and blue team scenarios.
  - Practical exercises based on emerging AI security challenges.
  - Live guidance and support from John Sotiropoulos.
  - Opportunities to explore and discuss real-world AI security incidents.
  
  Format & event details
  - Format: Live, online, interactive workshop
  - Style: Instructor-led demonstrations, guided exercises, red/blue team scenarios, and discussion
  - Date: 23rd January 2027
  - Time: 9:00 AM EST
  - Location: Online (joining link provided after registration)
  - Live online, real scenarios

This is a hands-on session rather than a traditional presentation. You’ll work through simulated scenarios, examine evidence, make decisions from both attacker and defender perspectives, and see how different security controls affect the outcome.

The workshop also provides a practical preview of themes explored in the forthcoming Second Edition of Adversarial AI by John Sotiropoulos.

🎟️ Reserve your seat now - Limited seats. Live support. Realistic Scenarios.