Course Objective & Scope
Letting a coding agent execute shell commands without constant human prompts unlocks massive productivity, yet running unconstrained on a developer's host machine risks leaking ~/.ssh and cloud credentials, wiping databases or executing malicious code via indirect prompt injection. This course teaches how to combine high agent autonomy with strong runtime isolation.
OS-level sandboxes (Seatbelt, Bubblewrap/Landlock), DevContainers and microVMs (gVisor, Firecracker), filesystem mounts and network egress allowlists, fine-grained MCP and command permission policies, pre-tool security hooks, and secret-injecting proxies are built and tested hands-on.
Who Should Attend?
- Platform, DevEx and infrastructure engineers
- Application and cloud security (AppSec / SecOps) engineers
- Senior engineers and architects rolling out autonomous coding agents across teams
- Prerequisite: familiarity with Linux/containers, networking fundamentals and CLI coding agents
Key Takeaways
- Isolate coding agents from the host OS, sibling repositories and personal credentials
- Eliminate permission-prompt fatigue safely by running autonomous modes inside hardened sandboxes
- Restrict network egress to approved package registries and APIs to block data exfiltration
- Author fine-grained command/MCP allowlists, denylists and deterministic pre-tool hooks
- Use credential-injecting proxies so raw secrets never enter the agent's environment or context window
Curriculum
Coding Agent Threat Model and Permission Fatigue
- Indirect prompt injection via untrusted repos, issues, logs and dependencies
- Why constant permission popups fail (approval fatigue) versus hard sandbox boundaries
Runtime Isolation: OS Sandboxes, DevContainers and MicroVMs
- Comparing lightweight OS sandboxes (Seatbelt, Bubblewrap, Landlock) with DevContainers
- High-isolation cloud execution with gVisor, Kata Containers and Firecracker microVMs
Filesystem Mounts, Network Egress and Resource Quotas
- Mounting only the workspace read-write while protecting
.git/hooks,~/.sshand cloud configs - Blocking HTTP/DNS data exfiltration with domain allowlist egress proxies
Tool Permissions, MCP Hardening and Pre-Tool Hooks
- Configuring permission profiles (
read-only,workspace-write) and command allowlists/denylists - Defending against MCP tool poisoning and blocking dangerous patterns with pre-tool hooks
Zero-Exposure Secret Management and Audit Logging
- Ephemeral credentials and boundary proxies that inject auth headers outside the sandbox
- Centralised audit logging of every shell command, file write and outbound connection
Other Courses in This Category
-
Scaling AI coding from solo tricks to team workflows: building version-controlled agent skills (SKILL.md), custom slash commands, and structured context handoffs across sessions, agents and engineering roles.
- Live Course
- 5 modules
-
Running headless coding agents inside GitHub Actions and GitLab CI pipelines for automated CI failure repair, issue-to-draft-PR workflows, dependency upgrades and strict security/cost guardrails.
- Live Course
- 5 modules
-
Practical training on using IDE coding assistants such as GitHub Copilot, Cursor and Windsurf effectively through context pinning, multi-file editing, test-backed verification and secure coding habits.
- Live Course
- 5 modules
-
Training on combining natural-language-driven software creation (vibe coding) in Cursor, Claude Code, Lovable and Bolt with modular architecture, Git checkpoints and security guardrails so prototypes stay maintainable.
- Live Course
- 5 modules
-
Hands-on training in autonomous multi-file software engineering with Claude Code, Cursor Agent and OpenAI Codex CLI, covering Plan-Act-Verify loops, MCP tool integration and parallel development with Git worktrees.
- Live Course
- 5 modules
-
End-to-end product-to-engineering workflow training that turns business requirements, PRDs, user stories and Figma designs into technical specs, task slices, acceptance criteria and verified code.
- Live Course
- 4 modules
Get in touch
