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

Foundations: Threat Models for Agentic AI

• Classifying agentic threats: misuse, privilege escalation, data leakage, and supply-chain vulnerabilities.
• Understanding adversary profiles and attacker capabilities relevant to autonomous agents.
• Mapping assets, defining trust boundaries, and identifying critical control points for agents.

Governance, Policy, and Risk Management

• Establishing governance frameworks for agentic systems, including roles, responsibilities, and approval gates.
• Designing policies for acceptable use, escalation protocols, data handling, and auditability.
• Addressing compliance requirements and collecting evidence for audits.

Non-Human Identity & Authentication for Agents

• Creating identities for agents using service accounts, JWTs, and short-lived credentials.
• Applying least-privilege access patterns and just-in-time credentialing strategies.
• Managing the identity lifecycle, including rotation, delegation, and revocation.

Access Controls, Secrets, and Data Protection

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

Observability, Auditing, and Incident Response

• Designing telemetry for agent behavior, including intent tracing, command logging, and provenance tracking.
• Integrating with SIEM tools, setting alerting thresholds, and ensuring forensic readiness.
• Developing runbooks and playbooks for managing agent-related incidents and containment.

Red-Teaming Agentic Systems

• Planning red-team exercises with defined scope, rules of engagement, and safe failover mechanisms.
• Employing adversarial techniques such as prompt injection, tool misuse, chain-of-thought manipulation, and API abuse.
• Executing controlled attacks to measure exposure and impact.

Hardening and Mitigations

• Implementing engineering controls like response throttles, capability gating, and sandboxing.
• Applying policy and orchestration controls, including approval flows, human-in-the-loop mechanisms, and governance hooks.
• Utilizing model and prompt-level defenses such as input validation, canonicalization, and output filtering.

Operationalizing Safe Agent Deployments

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

Capstone: Red-Team / Blue-Team Exercise

• Conducting a simulated red-team attack against a sandboxed agent environment.
• Defending, detecting, and remediating as the blue team using established controls and telemetry.
• Presenting findings, a remediation plan, and recommended policy updates.

Summary and Next Steps