ai-safe2-framework

AI SAFE² v3.1 Surface: Example Context: v3.1 Current

Framework Home Examples Index Cross-Pillar Governance AISM NEXUS Dashboard

Current framework context: AI SAFE² v3.1. This example may preserve historical component versions or earlier framework references where they describe when the implementation was created. For current conformance, use the v3.1 framework and applicable profile requirements.

# Langflow Sovereign Runtime ### AI SAFE2 v3.1 Defense Package for Langflow (Visual DAG Builder + Webhook + MCP + RAG) **Cyber Strategy Institute** · MIT License · Framework: AI SAFE² v3.1

Langflow is not a Python runtime framework you can wrap with a callback hook.

It’s a visual DAG builder where nodes execute internally. You cannot intercept execution from outside the DAG. The sovereign runtime sits as an external API proxy — scanning every request before it reaches Langflow’s HTTP endpoints — plus an inline DAG node that gates component outputs within the flow itself.

Eight surfaces. Two enforcement layers. Zero dependencies.


Why Langflow Requires a Different Architecture

Every other runtime in this series (LangChain, LangGraph, CrewAI, Make.com, Cursor) has a Python execution boundary you can intercept with a class wrapper or callback. Langflow’s DAG executes internally. There is no hook.

The AI SAFE2 Langflow Sovereign Runtime implements two enforcement layers:

Layer 1: External API Proxy — scans all requests before they reach Langflow’s HTTP endpoints. Deploy in your reverse proxy, API gateway, or application middleware.

Layer 2: Inline DAG Node — the safe2_guardian_component.py component drops directly into Langflow flows between data-fetching components and Agent nodes, intercepting the LF-COMP surface from within the DAG.


Eight Surfaces. All Confirmed From Live Docs.

Surface Confirmed From Risk Method
LF-WHK docs.langflow.org/webhook Async background run — no output gate; unauthenticated = flow owner scan_webhook_payload()
LF-RUN docs.langflow.org/api tweaks overrides ANY component field; default session = shared memory scan_run_request()
LF-GVAR docs.langflow.org/configuration-global-variables X-LANGFLOW-GLOBAL-VAR-LANGFLOW_DATABASE_URL redirects prod DB scan_global_var_headers()
LF-KNOW docs.langflow.org/knowledge RAG vector DB — poisoned doc persists for all future queries scan_knowledge_document()
LF-MCP docs.langflow.org/mcp-server AUTO-EXPOSE IS DEFAULT: all projects exposed instantly on create scan_mcp_config()
LF-FLOW docs.langflow.org/concepts-flows-import CustomComponent = raw Python via build engine; API key in JSON export scan_flow_json()
LF-INST docs.langflow.org/agents system_prompt saved in flow JSON; affects ALL future sessions scan_agent_instructions()
LF-COMP DAG architecture URL fetcher → injection → Agent node (multiple DAG hops) scan_component_output()

The Most Critical Surface: LF-MCP (Default-On)

Confirmed verbatim from Langflow’s documentation (June 2026):

“When you create a Langflow project, Langflow automatically adds the project to your MCP server’s configuration and makes the project’s flows available as MCP tools.”

The environment variable is LANGFLOW_ADD_PROJECTS_TO_MCP_SERVERS, which defaults to True.

Combined with LANGFLOW_AUTO_LOGIN=true (also a common default), you get an MCP server with no authentication exposing all your flows as tools to any connected MCP client. This is the highest-blast-radius default configuration in this series.

Fix before enabling MCP:

LANGFLOW_ADD_PROJECTS_TO_MCP_SERVERS=false
LANGFLOW_AUTO_LOGIN=false

The Proxy + DAG Architecture

External Client
      │
      ▼
┌─────────────────────────────────┐
│   AI SAFE2 API Proxy Layer      │
│   scan_webhook_payload()        │  ← LF-WHK, LF-RUN, LF-GVAR
│   scan_run_request()            │
│   scan_global_var_headers()     │
└──────────────┬──────────────────┘
               │ clean requests only
               ▼
┌─────────────────────────────────┐
│   Langflow Server (:7860)       │
│   POST /api/v1/webhook/$FLOW    │
│   POST /api/v1/run/$FLOW        │
└──────────────┬──────────────────┘
               │ DAG execution
               ▼
┌─────────────────────────────────┐
│   [URL Fetcher] → [AI SAFE2     │
│   Guardian] → [Agent Node]      │  ← LF-COMP (safe2_guardian_component.py)
└─────────────────────────────────┘

Package Contents

examples/langflow-sovereign-runtime/
│
├── enforcement/
│   ├── ai_safe2_engine.py          NEXUS kernel — stdlib only
│   ├── sovereign_langflow.py       8-surface Langflow enforcement class
│   └── __init__.py
│
├── langflow-component/
│   └── safe2_guardian_component.py Inline DAG node — drop into Langflow as CustomComponent
│
├── controls/
│   └── policy.yaml                 Machine-readable control registry
│
├── integrations/
│   ├── mcp-security.md             MCP safe configuration guide
│   └── langsmith-integration.md    AI SAFE2 + LangSmith observability
│
├── ci-cd/
│   └── github-actions-langflow-gate.yml
│
├── reports/                        Audit logs (gitignore)
├── smoke_test.py                   21/21 adversarial tests
├── requirements.txt
├── QUICKSTART.md
├── ANNOUNCEMENT.md
└── README.md

Quick Start

cd examples/langflow-sovereign-runtime
PYTHONPATH=enforcement python3 smoke_test.py
# TOTAL: 21/21 -- SOVEREIGN BASELINE VERIFIED

AI SAFE2 Pillar Coverage

Pillar Controls Langflow Enforcement
P1 Sanitize-Isolate P1.T1.2, P1.T1.5, P1.T1.9, P1.T1.10, P1.T1.4_ADV, P1.T2.5, S1.3, S1.5, S1.6 All 8 LF surfaces
P2 Audit-Inventory P2.T3.1, A2.5 SHA-256 JSONL per session
P3 Fail-Safe P3.T5.5 Ops rate limiting
P4 Engage-Monitor M4.5 Global var header monitoring; MCP tool monitoring
P5 Evolve-Educate E5.1 Love Equation + GREEN/YELLOW/RED
CP Cross-Pillar CP.4 MCP project allowlist governance

Known Enforcement Gaps

  1. DAG internal execution — Components between safe2_guardian nodes execute without enforcement. Place the guardian after every data-fetching component, not just at the first hop.
  2. Langflow UI direct actions — Knowledge uploads, flow imports, and system_prompt edits via the Langflow browser UI bypass the proxy layer. Pair with Langflow’s RBAC and SSO.
  3. Custom component validation depth — The flow JSON scanner validates code fields in template nodes but doesn’t execute the code. Complex obfuscated payloads may evade static pattern matching. Review all CustomComponent code manually before import.

Connect to the NEXUS Mesh

See the centralized Langflow legacy integration guide.

examples/
├── langflow-sovereign-runtime/    ← THIS PACKAGE
├── make-sovereign-runtime/
├── xai-grok-sovereign-runtime/
├── lovable-sovereign-runtime/
├── manus-sovereign-runtime/
└── cursor-sovereign-runtime/

MIT License — Cyber Strategy Institute “Engineered Certainty for the Agentic Age.”


Repository navigation

Examples Index Framework Home Cross-Pillar Governance NEXUS Scanner MCP Profile

AI SAFE² v3.1 | Cyber Strategy Institute