Workshops: Secure Agentic Coding & Governance

Sandboxing Coding Agents: Permissions & Least Privilege Workshop

A coding agent runs commands with the rights of the person who starts it. Without isolation it can delete files, read credentials or send data to the internet, whether by mistake or through manipulation. In this workshop your team builds a layered sandbox in which agents work productively with as few prompts as possible, but can only do what they actually need.

Who is this workshop for?

  • Security engineers and application security teams
  • Platform teams that provide developer environments
  • Tech leads who approve agents for their teams
  • Administrators of developer workstations and CI runners
  • Security architects

What you will take away

  • Assess which agent actions could cause which damage in your environment.
  • Configure permission modes and allow and deny rules in Claude Code, OpenAI Codex and other tools.
  • Build a hardened dev container for one of your repositories in which an agent may work autonomously.
  • Restrict outbound traffic with default-deny and an allowlist, and verify it with exfiltration tests.
  • Define how exceptions are requested, logged and reviewed.

Agenda

  1. Module 1

    Threat picture

    • Destructive commands, data exfiltration, lateral movement
    • Malicious install scripts and tampered dependencies
    • Why approval prompts alone do not protect: habituation and click fatigue
  2. Module 2

    Built-in tool controls

    • Permission modes and rules in Claude Code, approval and sandbox modes in OpenAI Codex
    • Command allowlists in Cursor and GitHub Copilot
    • Centrally managed settings for the whole team
  3. Module 3

    Isolation layers

    • Dev containers, virtual machines and operating-system sandboxing
    • Ephemeral cloud environments for asynchronous agents
    • Exercise: build a hardened dev container for your own repository
  4. Module 4

    Network egress

    • Default-deny with an allowlist for package registries and the model API
    • Proxies, DNS control and logging
    • Exercise: exfiltration attempts against your own sandbox
  5. Module 5

    Credentials and file system

    • Read-only mounts, no access to home directories and SSH keys
    • Short-lived, narrowly scoped tokens instead of personal credentials
    • Separate rights for local, CI and cloud agents
  6. Module 6

    Operations and governance

    • Policy as code and versioned sandbox templates
    • Exception process and logging
    • Involving the works council when developer activity is logged

Layers, not a single switch

No single mechanism is enough. Approval prompts get clicked away after the hundredth time, deny rules can be bypassed by detours, and a container boundary does not stop exfiltration if the network is open. So we combine tool rules, isolation, network control and minimal credentials.

The aim is a setup in which an agent can work autonomously because a mistake or a manipulation can only cause limited damage.

What you take away

You leave with concrete artefacts you can adopt in your repositories straight away.

  • A dev container template with hardened configuration
  • An egress allowlist for your typical package sources
  • Permission rules for the agents you use
  • A checklist for approving new agent setups

Experience from our own operations

Our AI Code Engineers work with coding agents in isolated environments every day. jocapps, the company behind neooutput, is ISO 27001 certified, and the measures shown fit well into an existing ISMS.

Frequently asked questions

Is a mode without approval prompts ever acceptable?

Yes, but only in a genuinely isolated environment without credentials and with restricted network access. That is exactly the environment we build during the workshop.

Containers or virtual machines?

Containers are fast and sufficient for many teams; virtual machines offer a stronger boundary. We discuss which option fits your risk profile and infrastructure.

Won't this slow developers down?

Done well, the opposite happens: fewer approval prompts and clear rules allow more autonomy. The one-off setup effort pays off across all repositories.

Workshop

Get in touch

How much autonomy can you safely give your coding agents?

Tell us roughly how many people will join and pick dates that work – we will get back to you within 24 hours with a proposal for content, format and schedule.

Prefer to talk first?

15 minutes, free and without obligation – with our AI Trainers and AI Code Engineers.

Book a call for free (opens in new tab) +49 761 519 751 50 (24/7)

Request a training / workshop

Workshop: Sandboxing Coding Agents: Permissions & Least Privilege Workshop · 1 day

Dates that work for you (optional)

Suggestions about 2 weeks ahead – pick as many as you like

Reply usually within 24 hours · jocapps is ISO 27001 certified · Privacy