Safe & Secure Agentic AI: Governance, Identity, and Red-Teaming Training Course

Foundations: Threat Models for Agentic AI

• Categorizing agentic threats: including misuse, privilege escalation, data leakage, and supply-chain risks.
• Defining adversary profiles and attacker capabilities uniquely associated with autonomous agents.
• Mapping critical assets, trust boundaries, and control points specific to agent operations.

Governance, Policy, and Risk Management

• Establishing governance frameworks for agentic systems, covering roles, responsibilities, and approval gates.
• Crafting policies for acceptable use, escalation rules, data handling, and ensuring auditability.
• Addressing compliance requirements and methods for collecting evidence for audits.

Non-Human Identity & Authentication for Agents

• Designing agent identities using service accounts, JWTs, and short-lived credentials.
• Implementing least-privilege access patterns and just-in-time credentialing strategies.
• Managing the identity lifecycle, including rotation, delegation, and revocation protocols.

Access Controls, Secrets, and Data Protection

• Applying fine-grained access control models and capability-based patterns for agent interactions.
• Securing secrets management, encryption in transit and at rest, and enforcing data minimization.
• Safeguarding sensitive knowledge sources and PII from unauthorized access by agents.

Observability, Auditing, and Incident Response

• Creating telemetry structures for agent behavior, including intent tracing, command logs, and provenance tracking.
• Integrating with SIEM systems, setting alerting thresholds, and ensuring forensic readiness.
• Developing runbooks and playbooks for responding to and containing agent-related incidents.

Red-Teaming Agentic Systems

• Planning red-team engagements: defining scope, rules of engagement, and safe failover mechanisms.
• Exploring adversarial techniques such as prompt injection, tool misuse, chain-of-thought manipulation, and API abuse.
• Executing controlled attacks to measure exposure and assess potential impact.

Hardening and Mitigations

• Implementing engineering controls like response throttles, capability gating, and sandboxing.
• Establishing policy and orchestration controls, including approval flows, human-in-the-loop mechanisms, and governance hooks.
• Deploying model and prompt-level defenses through input validation, canonicalization, and output filters.

Operationalizing Safe Agent Deployments

• Adopting deployment patterns such as staging, canary releases, and progressive rollouts for agents.
• Maintaining change control, robust testing pipelines, and pre-deployment safety checks.
• Fostering cross-functional governance among security, legal, product, and operations teams.

Capstone: Red-Team / Blue-Team Exercise

• Executing a simulated red-team attack against a sandboxed agent environment.
• Acting as the blue team to defend, detect, and remediate using established controls and telemetry.
• Presenting findings, outlining a remediation plan, and proposing policy updates.

Summary and Next Steps