01Module 09: Agent SafetyRisk taxonomy, minimal footprint, prompt injection defense, guardrails, human oversight, sandboxing, and responsible deployment.0201 - Agent Risk TaxonomyEight categories of agent risk, the confused deputy problem, severity matrices, and a Python risk assessment module.0302 - Minimal Footprint PrincipleLeast privilege, reversibility preference, scope confirmation, and a Python minimal-footprint agent wrapper.0403 - Prompt Injection in AgentsIndirect prompt injection attacks, real-world examples, detection and defense strategies, and a Python injection defense system.0504 - Guardrails and Action ValidationPre- and post-action guardrails, composable validators, denylist enforcement, rate limiting, and a complete Python guardrail pipeline.06Human Oversight MechanismsDesign human oversight that is meaningful, not performative - risk-based interruption, async approval queues, audit trails, and graduated autonomy.07Sandboxing Agent EnvironmentsContain the blast radius of any agent failure - process isolation, Docker security hardening, network policy, E2B cloud sandboxes, and escape vector prevention.08Responsible Agentic AISafety principles, EU AI Act compliance, accountability chains, bias, privacy, red-teaming, and building a safety review process for autonomous agent systems.
