| Framework Home | Cross-Pillar Governance | AISM | NEXUS | Dashboard |
| Previous: ← Scanner | Related: NEXUS | MCP Agent-to-Tool Profile |
Gateway action-risk routing scores are separate from organizational AISM maturity and the Technology Contribution Profile. Record provider, tool, and agent-to-agent boundaries separately. A mediated provider request does not establish containment of direct tool or delegated paths. Use the validation guide to test interception, authorization, actuation, outage, revocation, and legitimate utility before claiming runtime effectiveness.
The current gateway implementation is Gateway v3.0. AI SAFE² v3.1 did not silently rename this component.
“Core Gateway” is the north-south provider proxy. “NEXUS Sovereign Gateway” is the separate NEXUS component’s gateway scaffold and follows the NEXUS source version. These names and version numbers identify different components. See the component version map.
The gateway remains a compatible reference component for the north-south enforcement plane because the v3.1 release primarily changes protocol governance and the MCP agent-to-tool profile. Existing v3.0 gateway audit evidence therefore retains its v3.0 component identity until a separately tested gateway release changes that identity.
This distinction is intentional:
Client / Agent
|
v
HeartbeatMonitor hard stop on invalid liveness state
|
RateLimiter per-identity request limits
|
ProviderAdapter normalized provider boundary
|
RiskScorer action x sensitivity x history
|
HITLCircuitBreaker AUTO / MEDIUM / HIGH / CRITICAL
|
ImmutableAuditLog HMAC-SHA256 chained evidence
|
Provider Dispatch Anthropic / OpenAI / Gemini / Ollama / OpenRouter
|
ResponseScanner outbound exfiltration and injection checks
|
v
Client Response
| Component | Function |
|---|---|
HeartbeatMonitor |
Validates liveness state and activates safe mode on invalid conditions |
RateLimiter |
Bounds request volume per identity |
ProviderAdapter |
Normalizes provider requests while preserving the original upstream payload |
RiskScorer |
Computes the gateway risk vector used for HITL routing |
HITLCircuitBreaker |
Applies tiered human authorization requirements |
ImmutableAuditLog |
Produces HMAC-chained audit evidence and detects tampering |
ResponseScanner |
Inspects upstream output before return to the caller |
SafeMode |
Blocks normal traffic until operator-controlled recovery |
| Plane | Primary component/path | Gateway role |
|---|---|---|
| North-south | Gateway | Primary reference enforcement component |
| East-west | NEXUS | Gateway may carry/record NEXUS context but is not the primary A2A governance layer |
| Agent-to-tool | CP.5.MCP + NEXUS MCP adapter/toolkit | Gateway is not a substitute for MCP-specific authorization/profile enforcement |
The gateway therefore should not be used as evidence that MCP-14 through MCP-19 are implemented. Those controls must be evaluated on the actual agent-to-tool path.
python3 -m pip install fastapi uvicorn httpx pyyaml requests
export AUDIT_CHAIN_KEY="$(openssl rand -hex 32)"
export OPERATOR_DEACTIVATION_KEY="$(openssl rand -hex 16)"
Set the credential for the selected provider, for example:
export ANTHROPIC_API_KEY="..."
# export OPENAI_API_KEY="..."
# export GEMINI_API_KEY="..."
# export OPENROUTER_API_KEY="..."
uvicorn gateway.main:app --host 127.0.0.1 --port 8080
| Provider | Credential/config |
|---|---|
| Anthropic | ANTHROPIC_API_KEY |
| OpenAI | OPENAI_API_KEY |
| Gemini | GEMINI_API_KEY |
| Ollama | local endpoint, normally no API key |
| OpenRouter | OPENROUTER_API_KEY |
All providers pass through the same gateway policy stages; provider adapters handle upstream-specific transport details.
The gateway contains compatibility hooks for NEXUS context and can record NEXUS-related identity/delegation evidence when present.
That does not make the gateway the NEXUS implementation. NEXUS remains the CSI reference path for east-west governance and the v3.1 agent-to-tool adapter contract.
See NEXUS for the current implementation boundary.
Gateway audit records intentionally identify the gateway component version that created them. Existing Gateway v3.0 evidence should remain v3.0 evidence.
The audit-chain genesis anchor must not be changed merely because the framework version advances. Doing so would break verification of prior chained records.
A future Gateway v3.1 release should define and test an explicit evidence migration/versioning policy rather than silently rewriting the v3.0 anchor.
Safe mode blocks governed traffic until explicitly deactivated by the authorized operator path.
Typical triggers include:
The safety path remains outside ordinary model authority.
| Framework Home | Cross-Pillar Governance | AISM | NEXUS | Scanner | Dashboard |
AI SAFE² v3.1 framework · Gateway v3.0 component · Cyber Strategy Institute