Security and Privacy in Agentic Systems is a one-day, hands-on course. It is built for teams responsible for securing LLM-powered applications, autonomous agents, and AI workflows that access tools, data, identities, and external systems. Modern agentic systems introduce risks beyond traditional application security. Prompt injection can influence tool use, retrieved content can carry malicious instructions, and agents may act with excessive authority. Multi-step workflows can also hide unsafe delegation, while logs or embeddings can expose sensitive data.
The day opens by mapping the agentic AI threat landscape: LLM and agentic application risks, and prompt injection and jailbreak techniques. It also covers excessive agency, tool misuse, identity abuse, and confused deputy risks. A first lab has participants threat-model an agentic workflow and identify its key trust boundaries. Attendees then secure the tools, MCP servers, APIs, and actions an agent can take. They apply least-privilege tool design, input and output validation, scoped credentials, and approval flows. They then add sandboxing, egress limits, and guardrails to a tool-using workflow.
A data protection module examines sensitive data exposure through prompts, retrieved content, and outputs, along with classification, minimization, redaction, retention, and tenant isolation practices. This is reinforced by a lab reviewing an agent’s data flow for privacy risks. The course closes with testing, monitoring, and incident response. Students build evaluation harnesses and adversarial test cases, and instrument observability for prompts, tool calls, and policy decisions. They also draft an incident response playbook for unsafe actions or data exposure.
This course gives participants a current, practical treatment of agentic security and privacy. It draws on OWASP LLM and Agentic Application guidance, NIST AI RMF and the Generative AI Profile, and real-world production controls. They leave with a security and privacy review checklist suitable for production AI systems.
Who Should Attend
Security professionals, privacy engineers, AI/platform teams, SREs, developers, architects, managers
What Attendees Will Learn
Upon completing Security and Privacy in Agentic Systems, participants will be able to:
- Explain the current threat landscape for LLM applications and agentic AI systems
- Identify agent-specific risks such as excessive agency, tool misuse, identity abuse, and unsafe delegation
- Apply controls for prompt injection, data leakage, insecure outputs, RAG weaknesses, and supply chain exposure
- Design safer tool, MCP, API, memory, and multi-agent boundaries for agentic workflows
- Evaluate privacy risks around user data, retrieved content, logs, prompts, embeddings, and model outputs
- Implement practical monitoring, audit, human review, and incident response practices for agent systems
- Map agentic security practices to OWASP, NIST AI RMF, governance, and compliance expectations
Prerequisites
Participants should understand basic LLM and agent concepts, such as prompts, context, tools, retrieval, and model outputs. Prior security, privacy, cloud, platform, or application operations experience is helpful. Programming experience is useful for labs but not required.