In this briefing
  1. 01OpenAI Places File and Network Boundaries in an Operating-System Sandbox
  2. 02Microsoft Places Business Policy Before an Action
  3. 03Docker Governs the Execution Environment Separately from Tool Permissions
  4. 04More Connectors Expand the Actions That Policy Must Cover
  5. →What to watch next
  6. ↗Sources and verification
Key points
  1. OpenAI has published the filesystem, network-proxy and layered-configuration mechanisms of the Codex sandbox. The mechanism can limit the access available to a command's process tree, but OpenAI also identifies continuing risks from broad read access and subsequent execution outside the sandbox.
  2. Microsoft turns input, model calls, tool calls, output and shutdown into policy intervention points, and Microsoft's managed service can return allow, deny or transform decisions. The service remains in public preview.
  3. Docker separates the execution environment, Model Context Protocol tools and audit records into three governance layers, while GitHub adds eight connectors that can operate on content, pipelines, databases and websites. Enterprises need every tool path to pass through an enforceable permission policy.
Signal 01

OpenAI Places File and Network Boundaries in an Operating-System Sandbox

OpenAI published the design of the Codex sandbox on 18 November. Codex uses the operating-system mechanisms of macOS, Linux and Windows to constrain local commands; Codex does not move every task into one common virtual machine or container. Users can limit filesystem and network access through permission profiles, and child processes started by a command inherit the same restrictions. OpenAI lists three network modes: no network access, full direct access, and access restricted by domain through a proxy.

OpenAI also describes the design boundaries. The default `:workspace` profile permits writes inside the workspace and disables network access, but the profile retains broad filesystem read access, so a command may still read sensitive configuration outside the workspace. OpenAI also gives the example of an Agent modifying a local software package ignored by Git, after which an editor may execute the modified software outside the sandbox. A sandbox can reduce command permissions, but a sandbox cannot by itself replace model training, human approval or software-supply-chain controls.

What this may mean for enterprise adoption

For enterprise adoption, an enterprise needs to define explicit file, network and process permissions for each class of Agent task, and administrator-defined limits need to override personal configuration. The enterprise also needs to test ignored files, build caches, development tools and subsequent execution after a task ends because these objects may cross a sandbox boundary. A permission profile becomes a verifiable control only when the enterprise can prove its effective scope, inheritance and bypass paths.

Signal 02

Microsoft Places Business Policy Before an Action

Microsoft announced agent controls on 18 November. The managed Azure service is built on the open-source Agent Hooks and Agent Control Specification. Agent Hooks defines intervention points at startup, input, model calls, tool calls, output and shutdown, and the runtime pauses at those points to request a decision. Agent Control Specification allows safety classifiers, large-language-model judges, deterministic business rules or an external application programming interface to provide evidence before a policy engine returns a decision.

Microsoft says the service can return allow, deny or transform decisions, and each evaluation can return a receipt that identifies the policy and version. Microsoft also states explicitly that the Agent host remains responsible for enforcing the returned decision. The service is currently in public preview. The public material can confirm the interface direction and example flow, but the public material cannot prove that every framework, cloud environment and high-concurrency workload has already achieved the same stability.

What this may mean for enterprise adoption

For enterprise adoption, an enterprise needs to express transfer limits, external data destinations, sensitive tools and human-review conditions as independent policies instead of placing those restrictions only in a system prompt. The enterprise also needs to retain the policy version, evaluation input, returned decision, host-enforcement result and final business receipt. A policy service returning allow does not mean that the business action completed correctly, and a policy service returning deny also needs to connect to exception handling and human takeover.

Signal 03

Docker Governs the Execution Environment Separately from Tool Permissions

Docker described three components of Docker AI Governance on 18 November. Docker Sandboxes gives each Agent a microVM with a separate filesystem, network and Docker Engine, while Sandbox Policies lets an enterprise restrict filesystem and network access centrally. Model Context Protocol Tool Governance controls server registration, tool calls, resources, prompts and approval requirements through Docker MCP Gateway. When policy enforcement is active, requests without a matching allow rule are denied.

Docker explicitly retains a gateway boundary. Model Context Protocol policies cover only activity handled through Docker MCP Gateway; if an Agent connects directly to a remote server, only the sandbox network policy governs that connection. Docker's current audit records retain governance metadata such as the principal, action, target, decision and time, but the audit records do not retain prompt or response content. Current coverage mainly includes sandbox-policy decisions and sandbox-session events.

What this may mean for enterprise adoption

For enterprise adoption, an enterprise needs to verify separately where an Agent runs, which gateway the Agent uses to call a tool and which activities actually enter an audit record. An enterprise cannot assume that a tool implementation will avoid changing data merely because the tool carries a read-only annotation; the enterprise also needs to verify server behaviour, argument permissions and the business-system receipt. The enterprise also needs to prevent an Agent from bypassing the governed gateway to reach the same remote service directly.

Signal 04

More Connectors Expand the Actions That Policy Must Cover

GitHub added eight GitHub-verified Model Context Protocol connectors to GitHub Copilot on 18 November. The new services are Box, Buildkite, Calendly, ClickUp, Mixpanel, Supabase, Webflow and Wix. GitHub lists capabilities that cover searching and managing content, operating build pipelines, managing calendars, querying product analytics, checking database schemas and managing website content. The connectors therefore do not provide only read-only knowledge retrieval.

GitHub marks the connectors as a public-preview capability for every GitHub Copilot plan. Administrators for GitHub Copilot Business and GitHub Copilot Enterprise need to enable the `Editor preview features` policy. If an enterprise maintains an allowlist by connector, an administrator also needs to add each new connector to that list. Some external services also require a service-side administrator to enable access separately. The identity of a connector verified by GitHub cannot replace enterprise approval for each action and data scope.

What this may mean for enterprise adoption

For enterprise adoption, a larger connector inventory expands both the business value of an Agent and the range of actions that the Agent can perform. An enterprise needs to authorise each connector, server, tool, action, data object and environment separately, and the enterprise needs to include preview status, external-service administration and the enterprise allowlist in the same release check. If an enterprise validates only the connector source, the enterprise still cannot prove that every tool call complies with business permissions.

Verification

Sources and verification

  1. Inside the Codex sandboxOpenAI · 2026-11-18 · Official documentation
  2. The Trust Layer: Securing Agentic AI with Docker AI GovernanceDocker · 2026-11-18 · Official announcement
  3. Connect Supabase, Box, Calendly, and more to CopilotGitHub · 2026-11-18 · Official announcement

Golden Data has edited this briefing from the public materials listed above. The original sources govern facts and figures. The enterprise relevance sections are Golden Data editorial analysis and do not constitute an endorsement of any third-party product.

← Back to AI Daily Briefing