Generate Sandbox Policy

Generate YAML sandbox network policies and network middleware configuration from API documentation and natural-language user requirements.

Overview

This skill translates a user's plain-language policy intent into a valid sandbox policy. The amount of detail the user provides determines the granularity of the generated policy — from broad L4 or preset-based policies (just a host:port) up to fine-grained per-endpoint L7 rules (full API docs).

The output is a network_policies YAML block, an optional network_middlewares block, and optionally a full policy file that conforms to the sandbox policy schema.

Step 1: Gather Inputs

Determine the Detail Tier

The user's input falls into one of three tiers. Work with whatever the user provides — do not require a higher tier than needed.

Minimal Tier (host + intent only)

The user provides API endpoints and a broad intent. No API docs needed.

Examples:

• "Allow curl to hit api.github.com, read-only"
• "Give claude full access to api.anthropic.com"
• "Let /usr/bin/myapp talk to internal-svc:8080 but only for reading"

This is sufficient for:

• L4-only policies (allow all traffic to host:port, no HTTP inspection)
• Preset-based L7 policies (read-only, read-write, full on all paths)

For this tier, default to:

• access: read-only when the user says "read", "browse", "view", "query", "fetch"
• access: read-write when the user says "read-write", "create", "update" (but not "delete")
• access: full when the user says "full access", "everything", "unrestricted"
• L4-only when the user says "just allow it", "pass through", or "no inspection". Omit protocol for explicit-proxy clients. Use protocol: tcp only when the workload must use native DNS and direct socket calls, the endpoint has a valid DNS hostname, and the selected runtime support (currently Docker and Podman).

Moderate Tier (host + partial path knowledge)

The user knows some API paths but doesn't have full docs.

Examples:

• "Allow GET on /api/v1/models and POST on /api/v1/completions at integrate.api.nvidia.com"
• "Read-only on /repos/** at api.github.com, but also allow POST on /repos/*/issues"

Generate explicit rules for the known paths. If the user also wants broader access beyond the specific paths, combine with a catch-all rule or suggest a preset instead.

Full Tier (complete API docs)

The user provides full API documentation. Accepted formats:

From the API docs, build an endpoint inventory — a list of (method, path, description) tuples. Group them logically (e.g., by resource or tag). Then generate precise rules that allow only what the user's intent requires.

Policy Intent

Regardless of tier, extract (or infer) these from the user's description:

If the host and access level are clear but binaries are not specified, ask the user which binary or process will be making the requests. Suggest common defaults like /usr/bin/curl, /usr/local/bin/claude, etc.

Step 2: Refine Scope (Clarification Loop)

Before generating the policy, proactively ask clarifying questions to help the user scope the policy down as narrowly as possible. The goal is the most restrictive policy that still satisfies the user's needs.

Required Clarifications

Always ask about these if the user hasn't already specified them:

Scoping-Down Questions

Ask these when the user's intent is broad and more specificity is possible:

Auto-Discovery of API Docs for Well-Known Services

When the user mentions a recognizable API host but hasn't provided docs, and the current tier is Minimal, attempt to upgrade to Full by searching for the API documentation online.

When to trigger:

• The host is a well-known public API (e.g., api.github.com, api.anthropic.com, api.openai.com, integrate.api.nvidia.com, api.stripe.com, api.slack.com, api.gitlab.com)
• The user has NOT already provided API docs
• The user has NOT explicitly asked for a broad preset ("just read-only, nothing fancy")

How to do it:

1. Tell the user: "I can look up the REST API docs for [service] to help generate a more precise policy. Want me to do that?"
2. If the user agrees (or hasn't declined), search for the docs:
   • Use WebSearch with a query like "[service name] REST API documentation endpoints" or "[service name] OpenAPI spec"
   • Look for official documentation URLs in the results
3. Fetch the docs page with WebFetch and extract the endpoint inventory (method + path pairs)
4. Use the discovered endpoints to offer tighter scoping: "I found [N] endpoints in the [service] API. Based on your intent, I can narrow the policy to just [subset]. Want me to do that, or keep the broader preset?"

When to skip:

• The user explicitly asked for a broad preset or said "don't bother with docs"
• The API is internal, private, or not publicly documented
• The host is not recognizable as a well-known service
• A previous search attempt for this host returned no useful results

Graceful fallback: If the search doesn't return usable API docs (results are irrelevant, docs are behind authentication, the page is too large to parse), fall back to the current tier without stalling. Say: "I couldn't find usable API docs for [host], so I'll generate the policy using a [preset/L4] approach. You can always provide docs later to tighten it."

When the User Can't Narrow Further

If the user confirms the policy must stay broad (they don't know the paths, need genuinely broad access, etc.), accept it but flag the breadth. Do not block policy generation — just make sure the warnings are visible in the output (see Step 6).

Iteration

You may need to go back and forth a few times. Keep the loop tight:

1. Ask one batch of clarifying questions (group related questions together)
2. Update your understanding based on the answer
3. If the answer reveals further scoping opportunities, ask a follow-up
4. Stop when the user confirms the scope or says to proceed

Do not over-interrogate. If the user has given a clear, specific request, skip clarification and go straight to generation. Only ask when there is genuine ambiguity or an opportunity to meaningfully reduce the attack surface.

Step 3: Read the Policy Schema

Read the full policy schema reference:

Read docs/reference/policy-schema.mdx

Key sections to reference:

• Policy Schema Reference — top-level structure
• network_policies — rule structure
• NetworkEndpoint fields — host, port, protocol, tls, enforcement, access, rules, allowed_ips
• L7Rule / L7Allow — method + path matching
• Access Presets — read-only, read-write, full
• Private IP Access via allowed_ips — CIDR allowlist for private IP space
• Network Middleware - top-level middleware configs, ordering, host selection, and failure behavior
• Validation Rules — what combinations are valid/invalid

When middleware is requested, also read the full operational reference:

Read docs/extensibility/supervisor-middleware.mdx

Also read the architecture overview for enforcement context. The default policy is baked into the community base image (ghcr.io/nvidia/openshell-community/sandboxes/base:latest). For reference, consult:

Read architecture/security-policy.md

Step 4: Choose Policy Shape

Follow this decision tree based on the detail tier and user intent:

Is L7 inspection needed?
├─ No (user wants pass-through / "just allow it")
│   ├─ Explicit-proxy client → omit protocol
│   └─ Native DNS/socket client with a DNS hostname on a supported runtime → protocol: tcp
│
└─ Yes (user wants method/path control)
    │
    ├─ Does a preset match the intent exactly?
    │   ├─ Read-only (GET, HEAD, OPTIONS) → access: read-only
    │   ├─ Read-write (no DELETE)          → access: read-write
    │   └─ Everything                      → access: full
    │
    └─ No preset fits (specific paths, mixed broad+narrow, exclude certain paths)
        └─ Build explicit rules list
            └─ Requires either known paths from the user or full API docs

Principle: always choose the simplest representation that satisfies the intent. A preset is preferable to explicit rules when it covers the use case.

TLS Decision

Critical: protocol: rest on port 443 without tls: terminate will not work — the proxy cannot inspect encrypted traffic. Always set tls: terminate when combining port 443 with L7 rules.

Middleware Decision

Add network_middlewares only when the user asks to inspect, transform, redact, or independently authorize admitted HTTP requests or client WebSocket text messages. Middleware runs after network and L7 policy admission and before provider credential injection.

• Use openshell/regex without gateway registration for fixed-pattern redaction of UTF-8 HTTP request bodies or complete client-to-upstream WebSocket text messages.
• Use an operator-owned middleware name only when it is already registered under [[openshell.supervisor.middleware]] and reachable from both the gateway and sandbox supervisors.
• Confirm that a requested WebSocket implementation exposes a WEBSOCKET_MESSAGE/PRE_CREDENTIALS binding. openshell/regex exposes this binding. A host-matched HTTP-only implementation may inspect the upgrade GET but does not join the post-upgrade chain; messages pass and OpenShell emits binding_not_selected coverage regardless of on_error.
• WebSocket middleware runs for both ws:// and wss:// and receives complete client text messages only. Binary messages pass under both error modes and emit unsupported_message_type coverage for active stages. Upstream-to-client messages remain uninspected. Do not claim that V1 provides all-message WebSocket inspection.
• Treat fail_open on WebSocket as a session-scoped bypass: if the stage stream fails, OpenShell disables it for later messages on that connection and emits a state-change finding. Prefer fail_closed for required redaction or authorization.
• on_error governs failures after an advertised operation binding is selected. It does not apply to an unadvertised WebSocket binding or binary-message pass-through. An explicit HTTP, WebSocket preflight, or WebSocket message denial is authoritative under both fail_open and fail_closed.
• Default on_error to fail_closed. Use fail_open only when bypassing the stage preserves the user's stated security requirement.
• Assign unique order values across the complete policy. Lower values run first, and at most 10 configs may be selected.
• Match the narrowest destination hosts possible with endpoints.include; use exclude when a broad selector has trusted exceptions.
• Do not select fail-closed middleware for tls: skip endpoints because the supervisor cannot inspect that traffic.

Mapping Paths to Glob Patterns (when building explicit rules)

Only needed for the Moderate and Full tiers. Translate API path parameters to glob patterns:

Path matching uses the runtime glob engine. Both * and ** may cross / boundaries; ? matches one character, and bracket classes such as [0-9] and [!0] are supported. Prefer segment-shaped patterns such as /repos/*/issues for readability, but do not rely on * to stop at /.

Building the Explicit Rules List

For each allowed operation, create an allow entry:

rules:
  - allow:
      method: GET
      path: "/api/v1/models/*"
  - allow:
      method: POST
      path: "/api/v1/completions"

Use the most specific pattern that covers the intent. Prefer narrow globs over ** when the API structure is known.

Step 5: Generate the Policy

Output Format

Generate a complete network_policies entry. Use this template:

network_policies:
  <policy_key>:
    name: <policy_key>
    endpoints:
      - host: <api_host>
        port: <port>
        protocol: rest          # Required for L7 inspection
        tls: terminate          # Required for HTTPS + L7
        enforcement: enforce    # or audit
        # Use ONE of: access OR rules (never both)
        access: <preset>        # read-only | read-write | full
        # OR
        rules:
          - allow:
              method: <METHOD>
              path: "<glob_pattern>"
        # Optional: allow private IP destinations (CIDR or exact IP)
        # allowed_ips:
        #   - "10.0.5.0/24"
    binaries:
      - { path: <binary_path> }

When middleware is requested, add it as a separate top-level map rather than nesting it under a network policy:

network_middlewares:
  <config_key>:
    name: <human_readable_name>
    middleware: <built_in_or_registered_name>
    order: 10
    config: {}
    on_error: fail_closed
    endpoints:
      include: ["<api_host>"]
      # exclude: ["<trusted_host>"]

The map key is the stable policy-local identity. Middleware selection is independent of the network policy entry that admitted the request.

Deny Rules

Use deny_rules to block specific dangerous operations while allowing broad access. Deny rules are evaluated after allow rules and take precedence. This is the inverse of the rules approach — instead of enumerating every allowed operation, you grant broad access and block a small set of dangerous ones.

# Example: Allow full access to GitHub but block admin operations
github_api:
  name: github_api
  endpoints:
    - host: api.github.com
      port: 443
      protocol: rest
      enforcement: enforce
      access: read-write
      deny_rules:
        - method: POST
          path: "/repos/*/pulls/*/reviews"
        - method: PUT
          path: "/repos/*/branches/*/protection"
        - method: "*"
          path: "/repos/*/rulesets"
  binaries:
    - { path: /usr/bin/curl }

Deny rules support the same matching capabilities as allow rules: method, path, command (SQL), and query parameter matchers. When generating policies, prefer deny rules when the user needs broad access with a small set of blocked operations — it produces a shorter, more maintainable policy than enumerating 60+ allow rules.

Private IP Destinations

When the endpoint resolves to a private IP (RFC 1918), the proxy's SSRF protection blocks the connection by default. Use allowed_ips to selectively allow specific private IP ranges:

• Host + allowlist: host + allowed_ips — domain must resolve to an IP in the allowlist
• Hostless allowlist: allowed_ips only (no host) — any domain on the port is allowed if it resolves to an IP in the allowlist

Loopback (127.0.0.0/8) and link-local (169.254.0.0/16) are always blocked regardless of allowed_ips.

# Example: Allow access to internal service at a known private IP range
internal_api:
  name: internal_api
  endpoints:
    - host: api.internal.corp
      port: 8080
      allowed_ips:
        - "10.0.5.0/24"
  binaries:
    - { path: /usr/bin/curl }

Policy Key Naming

Use descriptive snake_case keys: github_api, nvidia_inference, internal_service_readonly.

Multiple Endpoints

If the user needs access to multiple hosts or the same host with different rules, either:

1. Same policy — multiple entries in endpoints if the binary set is the same
2. Separate policies — different policy keys if the binary sets differ

Step 6: Validate and Warn

Before presenting the policy to the user, verify correctness and flag breadth concerns.

Hard Errors (would block sandbox startup)

• rules and access are NOT both present on the same endpoint
• If an L7 protocol is set, either rules or access is also present; protocol: tcp is L4-only and must not contain either field
• Every protocol: tcp endpoint has a valid DNS hostname; it is not hostless, an IP literal, a trailing-dot name, or a malformed DNS selector
• If tls: terminate is set, protocol is also set
• rules list is not empty when present
• If protocol: sql, enforcement is not enforce
• Every middleware config has a non-empty middleware name and non-empty endpoints.include
• Middleware order values are unique and no selected chain exceeds 10 stages
• No fail-closed middleware selector can cover a tls: skip endpoint
• Any required WebSocket control advertises WEBSOCKET_MESSAGE/PRE_CREDENTIALS, and the user understands that V1 does not inspect binary messages
• Endpoints contributed by a credentialed provider are not L4-only or tls: skip unless allow_uninspected_credentials: true explicitly records the exception

Schema Warnings (log-only, but should be fixed)

• protocol: rest on port 443 should have tls: terminate
• HTTP methods are standard: GET, HEAD, POST, PUT, DELETE, PATCH, OPTIONS, or *
• Credentialed destinations are also covered by the attached provider profile endpoint; policy admission alone does not authorize credential resolution

Structural Checks

• Every policy has name, endpoints, and binaries
• Every endpoint has host and port
• Every binary has path
• Policy key matches name field
• Every middleware selector has at most 32 combined include and exclude patterns

Breadth Warnings

Evaluate the generated policy for overly broad access and include warnings in the output to the user. These do not block generation, but the user must see them.

Format breadth warnings clearly in the output, e.g.:

⚠️ Breadth warning: This policy uses `access: full`, which allows all HTTP
methods (including DELETE) on all paths. If you don't need DELETE, consider
using `read-write` instead.

If there are no breadth warnings, say so explicitly: "No breadth concerns — this policy is well-scoped."

Step 7: Determine Output Mode

The policy needs to go somewhere. Determine which mode applies:

Mode A: Update an Existing Policy File

1. Read the existing file to understand current state:

   • What policies already exist under network_policies
   • What the filesystem_policy, landlock, and process sections look like
   • Whether the file uses compact ({ host: ..., port: ... }) or expanded YAML style

2. Check for conflicts:

   • Does a policy with the same key already exist? If so, ask the user whether to replace it, merge new endpoints/binaries into it, or use a different key.
   • Does an existing endpoint selector overlap the new selector? Compatible overlaps are allowed and can intentionally aggregate allow and deny rules. Reject or revise equally specific overlaps that disagree on connection or request-processing metadata, including TLS, destination constraints, protocol/parser behavior, enforcement, or credential handling. A more-specific path selector may override broader request-processing metadata.
   • If the sandbox uses an attached provider credential, confirm the provider profile also declares the intended host, port, and path. A sandbox policy allow cannot expand the profile's static credential binding.
   • For credential_signing, confirm an attached endpoint-bearing profile declares AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY and covers the signed endpoint. For an endpointless AWS profile, add credential_binding.provider with the exact attached provider name.

3. Apply the change:

   • Adding a new policy: Insert the new policy block under network_policies, maintaining the file's existing indentation and style.
   • Modifying an existing policy: Edit the specific policy in place — add/remove endpoints, change access presets, update rules, add binaries, etc. A rule authorizes every binary it lists to reach every endpoint and port it lists, so adding one binary grants it all of that rule's endpoints, and adding one endpoint grants it to all of that rule's binaries. State the resulting pairs to the user before writing them. When the user wants a binary to reach only part of a rule's endpoints, put that binary and those endpoints in a separate rule instead of extending the existing one. An empty binaries list means any binary, so leaving it off widens the rule to every process.
   • Removing a policy: Delete the policy block if the user asks.

4. Preserve everything else: Do not modify filesystem_policy, landlock, process, or other policies unless the user explicitly asks.

Mode B: Create a New Policy File

Generate a complete, standalone policy file. Use the full schema scaffolding:

version: 1

filesystem_policy:
  include_workdir: true
  read_only:
    - /usr
    - /lib
    - /proc
    - /dev/urandom
    - /app
    - /etc
    - /var/log
  read_write:
    - /tmp
    - /dev/null

landlock:
  compatibility: best_effort

network_policies:
  # <generated policies go here>

The filesystem_policy and landlock sections above are sensible defaults. Process identity is omitted so the selected compute driver can choose it. For Docker and Podman, each omitted identity field falls back to the image's OCI USER. Tell the user these are defaults and may need adjustment for their environment. Gateway inference is configured separately through openshell inference set/get. The generated network_policies block is the primary output.

When the user explicitly requests process.run_as_user or process.run_as_group, accept sandbox or a numeric UID/GID from 1 through 4294967294. Reject root (0) and the invalid identity sentinel (4294967295). Warn that a low numeric identity inherits permissions granted to the same ID on image files, mounted volumes, or devices.

If the user provides a file path, write to it. Otherwise, ask where to place it. A common convention is a project-local policy file (e.g., sandbox-policy.yaml) passed to openshell sandbox create --policy <path> or set via the OPENSHELL_SANDBOX_POLICY env var.

Mode C: Present Only (no file write)

Show the generated policy YAML with:

1. Summary — what the policy allows and denies, in plain language
2. The YAML — the complete network_policies block, ready to paste
3. Integration guidance:
   • Save to a local file and pass via openshell sandbox create --policy <path> or set OPENSHELL_SANDBOX_POLICY=<path>
   • For production: configure via the gateway
4. Caveats — any assumptions made, anything the user should verify

Step 8: Confirm and Refine

After presenting or applying the policy, ask if the user wants to:

• Tighten or loosen any rules
• Add more endpoints or binaries
• Switch between enforce/audit mode
• Move from a preset to explicit rules (or vice versa)
• Apply the policy to a file (if presented only)
• Create additional policies for other APIs

Quick Reference: Common Patterns

L4-Only (no HTTP inspection)

my_api:
  name: my_api
  endpoints:
    - { host: api.example.com, port: 443 }
  binaries:
    - { path: /usr/bin/curl }

HTTPS API with Read-Only Preset

my_api_readonly:
  name: my_api_readonly
  endpoints:
    - host: api.example.com
      port: 443
      protocol: rest
      tls: terminate
      enforcement: enforce
      access: read-only
  binaries:
    - { path: /usr/bin/curl }

HTTPS API with Explicit Rules

my_api_custom:
  name: my_api_custom
  endpoints:
    - host: api.example.com
      port: 443
      protocol: rest
      tls: terminate
      enforcement: enforce
      rules:
        - allow:
            method: GET
            path: "/api/v1/**"
        - allow:
            method: POST
            path: "/api/v1/data"
  binaries:
    - { path: /usr/bin/curl }
    - { path: /usr/local/bin/myapp }

HTTP (non-TLS) Internal API

internal_svc:
  name: internal_svc
  endpoints:
    - host: api.internal.svc
      port: 8080
      protocol: rest
      enforcement: enforce
      rules:
        - allow:
            method: GET
            path: "/health"
        - allow:
            method: POST
            path: "/api/v1/jobs"
  binaries:
    - { path: /usr/bin/curl }

Private IP Access (Host + Allowlist)

internal_db:
  name: internal_db
  endpoints:
    - host: db.internal.corp
      port: 5432
      allowed_ips:
        - "10.0.5.0/24"
  binaries:
    - { path: /usr/bin/curl }

Private IP Access (Hostless — Any Domain in Range)

private_services:
  name: private_services
  endpoints:
    - port: 8080
      allowed_ips:
        - "10.0.5.0/24"
        - "10.0.6.0/24"
  binaries:
    - { path: /usr/bin/curl }

Additional Resources

• Full policy schema: docs/reference/policy-schema.mdx
• Enforcement overview: architecture/security-policy.md
• Default policy: baked into the community base image (ghcr.io/nvidia/openshell-community/sandboxes/base:latest)
• Rego evaluation rules: sandbox-policy.rego
• For translation examples from real API docs, see examples.md