Der Shop ist online

Wir bauen Workstation AI-Maschinen für Websites, KI-Agenten und LLM-Training – online konfigurieren und kaufen.

Händlerpreise erhalten
KI-Produkte
KI-LabsOpenAI-AgentenClaude-AgentenGrok-BotWorkstation AI-natives CRMMarketingAlle Produkte
KI-Lösungen
Angewandte KIKI-AgentenGenerative KIKI-AutomatisierungKI-WeiterbildungKI-Pakete für KMUEnterprise AI LabKI-WorkstationsNVIDIA-Hardware & -SoftwareNVIDIA Accelerated ComputingPrivate KIGPU-Private-CloudGPU-ClusterEdge-KIKI nach Branche
KI-Services
PlattformmodernisierungDigital EngineeringDatengrundlagen & KIAutonomer BetriebKI-BeratungDevOps-AutomatisierungCybersicherheitSoftwareentwicklungAgentenentwicklungMLOps-EinrichtungGPU-LeasingGPU-Private-Cloud
Über uns
UnternehmenUnsere GeschichtePartnerKundengeschichten
KI-Rechenzentren
KI-RechenzentrenGPU-Private-CloudGPU-ClusterGPU-Leasing
KI-EinblickeJev-ReiheBenchmark LabModellauswahl
Technische Hinweise
Dokumentation
Linux-BetriebssystemeUbuntuRed Hat Enterprise LinuxSUSE LinuxAlpine LinuxDebian, Rocky & AlmaLinux
KI-Cloud-PlattformenAWS & Amazon BedrockGoogle Cloud & Vertex AIMicrosoft Azure & AI FoundryGPU-Clouds & Marktplätze
Private KI-RechenzentrenFür AnwaltskanzleienFür Steuer- und WirtschaftsprüfungskanzleienOn-Premises-PlattformenVollständig abgesicherte Cloud
ToolsRing PromoterJobshout
News
Blog
KontaktAnmelden

KI-Workstations, Multi-Agenten-KI-Software, GPU-Infrastruktur und intelligente Agentenlösungen für moderne Unternehmen.

Kontakt

KI-Lösungen

Angewandte KIKI-AgentenGenerative KIKI-AutomatisierungKI-WeiterbildungKI-Pakete für KMUEnterprise AI LabKI nach Branche

Produkte

Alle ProdukteWorkstation AI-natives CRM & ERPMarketingOpenAI-AgentenRing PromoterJobshout

Unternehmen

Über unsUnternehmenKI-RechenzentrenWarum WorkstationPartnerKundengeschichtenPreiseKontaktTenthMatrix

Ressourcen

KI-EinblickeTechnische HinweiseDokumentationLinux-BetriebssystemeKI-Cloud-PlattformenPrivate KI-RechenzentrenBlogSucheSitemap
Büro Belgien
Workstation SRL, Rue Vanderkindere 34, 1180 Uccle, BrusselsBE 0751.518.683Mo – Fr: 9:00 – 18:00 Uhr MEZ
+32 492 45 67 46
Hauptsitz Großbritannien
77-79 Marlowes, Hemel Hempstead HP1 1LFAnfahrt – Ausfahrt 20 der M25 (Outer London)Handelsregisternr.: 11641870Mo – Fr: 9:00 – 18:00 Uhr GMT
+44 7515 356 146
Büro Indien
#159 Sector 9, Pocket 1, DDA Flats, 110077 Dwarka, New Delhi
+91 98881 98841
Büro Taiwan
3F, No. 2-4, Aisi Rd., Ren'ai Dist., Keelung City 200, Taiwan基隆市仁愛區愛四路2之4號3樓Mo – Fr: 9:00 – 18:00 Uhr Taiwan-Zeit (GMT+8)
+886 926 892 701+32 492 45 67 46
Büro Estland
Aleksander Puškini tn 30-25, 20306 Narva, EstoniaMo - Fr: 9:00 - 18:00 Uhr estnischer Zeit (EET)
+32 492 45 67 46

© 2026 Workstation AI. Alle Rechte vorbehalten.

DatenschutzDatenlöschungCookiesNutzungsbedingungenWebsite-Sitemap

NVIDIA, das NVIDIA-Logo, RTX, RTX PRO, Blackwell, CUDA, NVLink, MGX, Omniverse, Spectrum-X, Quantum, BlueField, DGX und HGX sind Marken und/oder eingetragene Marken der NVIDIA Corporation in den USA und anderen Ländern. AMD, das AMD-Arrow-Logo, EPYC, Instinct, Ryzen und Threadripper sind Marken von Advanced Micro Devices, Inc. Dell und PowerEdge sind Marken der Dell Inc. oder ihrer Tochtergesellschaften. Intel und Xeon sind Marken der Intel Corporation oder ihrer Tochtergesellschaften. AMAX, AceleMax und LiquidMax sind Marken der AMAX Engineering Corporation. Alle anderen Produktnamen, Logos und Marken sind Eigentum ihrer jeweiligen Inhaber. Wir verwenden sie ausschließlich, um die Produkte zu kennzeichnen, die wir beschreiben, liefern oder unterstützen; ihre Verwendung bedeutet nicht, dass ihre Inhaber uns empfehlen.

Startseite / Technische Hinweise / Technology
AISecurityDevOps

Agentic AI Security: MCP OAuth, VPN & HashiCorp Vault Leases

Enterprise reference: MCP OAuth 2.1 + EMA, VPN integration, Vault dynamic secrets, auto-rotation, lease revoke, and vendor setup matrix

8. August 2026Technology7 min read

This Workstation technical brief covers agentic AI security and authentication: securing enterprise agents that call tools through MCP (Model Context Protocol) with OAuth 2.1, placing high-risk MCP endpoints behind VPN / private networking, and managing tool credentials with HashiCorp Vault — dynamic secrets, leases, auto-rotation, and immediate revoke. It recommends industry-aligned setups for Claude, OpenAI, and Cursor.

Agentic AI security: MCP OAuth, VPN, Vault leases

Navigation. Companion blog: short digest. Related: deep learning systems, Bedrock agents, multi-agent teams.
Agent digest (machine-readable).
  • Problem: Agents with static long-lived tool keys + public MCP = credential sprawl and lateral movement.
  • Auth: Remote MCP = OAuth 2.1 resource server + RFC 9728 PRM + RFC 8707 audience; PKCE mandatory.
  • Enterprise: Prefer MCP Enterprise-Managed Authorization (EMA / ID-JAG) via corporate IdP.
  • Network: Internal MCP and Vault on VPN/private link; egress allowlists per tool.
  • Secrets: Vault dynamic secrets + lease TTL; auto-rotate static roles; revoke on session end.
  • Split: LLM provider API keys != MCP access tokens != backend tool secrets.

Sources: MCP Authorization, Vault leases, Vault AI agent validated pattern, Claude connector auth. Verify current vendor docs before production rollout.

1. Threat model for enterprise agents

An agent is a privileged automation principal. Typical failure modes:

  • Secret sprawl — API keys in MCP JSON, .env committed to git, or pasted into prompts.
  • Token substitution — access token minted for server A accepted by server B (missing audience binding).
  • Over-scoped tools — one MCP can write prod DBs and open firewalls with the same session.
  • Unaudited OBO — agent acts without linking actions to a human identity.
  • Public MCP — internal tools reachable from the internet without VPN or private link.

Controls must cover identity (who), authorization (what tools), secrets (with what credentials), and network (from where).

2. MCP authentication: industry standard

For HTTP-based remote MCP, the specification aligns with OAuth 2.1:

  • MCP server = OAuth resource server; clients send Authorization: Bearer.
  • PKCE is mandatory; implicit grant is prohibited.
  • Servers MUST publish Protected Resource Metadata (RFC 9728) so clients discover the authorization server.
  • Use Resource Indicators (RFC 8707) so tokens are audience-bound to that MCP server.
  • Authorization Server Metadata (RFC 8414) and/or OIDC discovery for AS capabilities.
  • Dynamic Client Registration (RFC 7591) is recommended where clients must onboard without manual client IDs.
# Protected Resource Metadata (conceptual)
{
  "resource": "https://mcp.internal.example/mcp",
  "authorization_servers": ["https://auth.example.com"],
  "scopes_supported": ["mcp:tools", "mcp:resources"]
}

STDIO / local MCP is different: prefer environment credentials injected by Vault Agent — do not force browser OAuth on every desktop tool. Remote/public MCP should implement OAuth 2.1.

2.1 Enterprise-Managed Authorization (zero-touch)

The MCP Enterprise-Managed Authorization extension lets the corporate IdP (Okta, Entra ID, etc.) grant access to approved MCP servers at SSO using an ID-JAG (Identity Assertion JWT Authorization Grant), avoiding per-server consent fatigue. Adopt EMA for org-wide Claude / IDE / agent fleets.

3. VPN and private MCP integration

OAuth authenticates the client; it does not replace network isolation.

  • Host internal MCP servers and Vault on private CIDR (VPN, PrivateLink, Tailscale/WireGuard mesh, or service mesh mTLS).
  • Bind MCP listeners to private interfaces; block public 443 unless the product is intentionally internet-facing.
  • Apply egress allowlists from the agent runtime: only approved APIs/MCP hosts.
  • Split planes: developer laptops on corp VPN for Cursor; server-side agents in VPC with no internet MCP except approved SaaS.

Pattern: VPN for reachability + OAuth for authorization + Vault for secrets.

Figure A: IdP MCP Vault VPN architecture

4. HashiCorp Vault for agent secrets

Static long-lived tool keys are incompatible with agent blast radius. Vault provides:

4.1 Dynamic secrets + leases

Every dynamic secret returns a lease_id and TTL. The consumer must renew (if allowed) or request a replacement before expiry. When the lease ends, Vault can revoke the credential at the provider. This forces check-in, improves audit logs, and shrinks exposure windows.

4.2 Auto-rotation

For static roles (e.g. database password with rotation_period), Vault rotates on a schedule. Vault Agent templates re-fetch near end of life (default lease_renewal_threshold ~0.9 of TTL) and can restart a child process when credentials change.

4.3 Recommended TTLs for agents

Risk class Example TTL guidance
Critical writeProd DB mutate, IAM admin5-15 minutes; revoke on tool end
Read / stagingRead replicas, ticket APIs30-60 minutes
LLM provider keyOpenAI / Anthropic org keyVault-managed; rotate on schedule; never in MCP JSON

Figure B: Vault lease lifecycle

4.4 Vault + user attribution (validated pattern)

HashiCorp’s validated pattern: user authenticates; agent receives an on-behalf-of (OBO) token; tools authenticate to Vault with JWT; Vault maps claims to policies and issues scoped dynamic secrets. Audit trails link secret issuance to the human, not a shared robot account.

Vault Enterprise adds Agent Registry and OAuth resource server profiles so enrolled agents present OAuth JWTs without a separate Vault login step — with agent-specific constraints for delegation / OBO.

# Conceptual agent tool hook (do not ship secrets to the model)
vault_token = login_jwt(obo_token)           # Vault auth
secret = vault.read("database/creds/agent-ro")
lease_id, ttl = secret["lease_id"], secret["lease_duration"]
try:
    run_tool(db_url=secret["data"])          # use within TTL
finally:
    vault.lease.revoke(lease_id)             # or let TTL expire

5. Recommended setups by vendor

Figure C: Claude OpenAI Cursor patterns

5.1 Claude (Anthropic)

  • Prefer OAuth for remote MCP connectors; return 401 with WWW-Authenticate pointing at PRM so clients can discover auth.
  • Never put tokens/API keys in connector URL query strings (logged, cached, prohibited by MCP token rules).
  • Claude Code: local OAuth with secure token store and refresh; plugins must not read tokens.
  • Hosted Claude: Anthropic-managed client credentials for consenting users; still keep tool secrets in Vault behind your MCP.
  • Enterprise: align IdP with MCP EMA for zero-touch server access.

5.2 OpenAI (Agents / tools)

  • Separate model API keys from tool credentials; different rotation and blast radius.
  • Run agent runtimes in VPC; call private MCP over private networking.
  • Pass user-attributed tokens into tools; authenticate to Vault with JWT; mint dynamic secrets per invocation.
  • Log every tool call with user + agent + lease_id for compliance.

5.3 Cursor

  • MCP server config: reference environment variables only — never hardcode secrets in mcp.json.
  • STDIO servers: run under Vault Agent (template or env) so leases rotate without developers copying passwords.
  • Remote MCP: OAuth when the server supports it; otherwise corp VPN + short-lived bearer from Vault.
  • Team policy: allowlist approved MCP servers; block untrusted community MCPs on prod codebases.
  • Keep .env / secret files out of context via ignore rules; project rules: forbid pasting secrets into chat.

6. Reference control matrix

Layer Control Anti-pattern
IdentitySSO IdP + EMA / OAuth PKCEShared robot password
MCPPRM + audience-bound tokensTokens in URL; no expiry
NetworkVPN / PrivateLink + egress ACLPublic MCP for prod tools
SecretsVault lease + rotate + revokeYear-long keys in mcp.json
OpsAudit IdP+MCP+Vault; human gatesUngated prod deploys / payments

7. Implementation checklist

  1. Inventory every MCP server and classify public vs private.
  2. Implement OAuth 2.1 + PRM on all remote MCP; enable EMA with corporate IdP.
  3. Move MCP + Vault behind VPN/private networking; document how Cursor/Claude join the network.
  4. Replace static tool keys with Vault dynamic secrets; set TTLs by risk class.
  5. Wire Vault Agent (or SDK renew/revoke) into agent runtimes; revoke leases on session end.
  6. Separate LLM provider keys; store and rotate them in Vault — never in prompts or MCP config.
  7. Add human approval gates for money, identity changes, and production deploys.
  8. Test: expired lease fails closed; wrong-audience token rejected; public path blocked.
Workstation routing rule. Front door = MCP OAuth (who may call tools). Network door = VPN/mesh (from where). Secret door = Vault leases (with what). If any door is propped open with a static key, treat it as an incident waiting to happen.
Related reading. Edge WAF on the same live-config OpenResty proxy: WSL Proxy WAF blog · explainable edge WAF article · product overview · wslproxy.org/waf · tour youtu.be/r10XSonA5JE.

Published by Workstation — enterprise automation, multi-agent platforms, and secure delivery on Kubernetes.

Diesen Artikel teilen

Mehr zu Technology

Laptops für KI-Entwickler: LLMs lokal unter Linux mit GPUs mit 16 GB und 24 GB betreiben (Leitfaden 2026)

Laptops für KI-Entwickler: LLMs lokal unter Linux mit GPUs mit 16 GB und 24 GB betreiben (Leitfaden 2026)

Was auf eine RTX 5080 oder RTX 5090 Laptop GPU passt, die Speicherrechnung, die lohnenden offenen Modelle, die Einrichtung von Ubuntu 26.04 mit dem NVIDIA-Stack, Ollama, llama.cpp und vLLM, Coding-Assistenten, LoRA-Fine-Tuning, Sicherheit und wann Arbeit auf einen Server gehört

Weiterlesen
ML- und LLM-Systeme testen, deren Korrektheit sich nicht beweisen lässt: ein Python-Evidenz-Kit

ML- und LLM-Systeme testen, deren Korrektheit sich nicht beweisen lässt: ein Python-Evidenz-Kit

Evidenz statt Beweise, in etwa 400 Zeilen Python mit echter Ausgabe: Slice-Metriken mit Bootstrap-Intervallen und einem Pass/Fail-Vertrag, Property- und metamorphe Tests mit Hypothesis, Kalibrierung und Conformal Prediction, Pydantic-Verträge und Regressionsevaluierungen für LLM-Ausgaben, Drift-Prüfungen, ein CI-Gate und die exakte Verifikation eines kleinen neuronalen Netzes

Weiterlesen
TorchCurves in der Praxis: lernbare B-Spline- und Legendre-Kurven in PyTorch

TorchCurves in der Praxis: lernbare B-Spline- und Legendre-Kurven in PyTorch

torchcurves installieren, eine Kurve anpassen und dimensionieren, B-Spline oder Legendre und die Eingabeabbildung wählen, eine monotone Kalibrierungskurve und ein additives Modell auf tabellarischen Daten bauen, mit SciPy, scikit-learn, einem KAN und einem MLP vergleichen, Koeffizienten nach JSON exportieren, testen und in der Produktion betreiben, mit echten Ausgaben

Weiterlesen