The independent verification and accountability layer for consequential AI-agent actions.
Runtime governance complements your existing stack — Nexus proves what agents actually did to systems of record.
Nexus combines runtime action governance, trajectory-aware control, and tamper-evident SHA-256 Universal Action Receipts (UAR) in a local-first, air-gapped deployment layer.
SEE → CONTROL → VERIFY → PROVE
P99 runtime intercept: 6.1ms (Nexus benchmark harness)
Execution → state change → cryptographic proof
Intercept tool calls, enforce policy at the API boundary, and seal Universal Action Receipts — before/after state hashes and SHA-256 evidence auditors can reproduce.
Platform Nav Preview — Live Dashboard Header
nex_••••••••••••4421Telemetry Active live signal · API Key badge · Action Control Plane shortcuts
Accountability layer
SEE → CONTROL → VERIFY → PROVE
Open harness
core-ai-firewall/harness
P99 6.1ms
measured in Nexus benchmark harness test environment
UAR receipts
Tamper-evident SHA-256
Outcome verification
Watch the cinematic narrative for Financial Deception, Data Exfiltration, and Privilege Escalation — then run the sandbox to verify Ed25519 / SHA-256 action proofs and export UAR v2.
Tripartite outcomes · VERIFIED · BLOCKED · UNVERIFIED (ghost action)
Outcome verifier catches ledger mismatch · circuit breaker contains blast radius.
Unverified · Ghost Action: Agent reports HTTP 200 OK, but ledger/state cross-check finds no mutation (false success).
00s / 20s
Timeline
POST /v1/tools/process_refund · target: customer_acme · scope: payments:write
$ agent-finance-04 → nexus-shield-gateway intent: PROCESS_REFUND · amount: 50000 USD passport: pas_live_88192a · delegation_depth: 1
Unverified · Ghost Action: Agent reports HTTP 200 OK, but ledger/state cross-check finds no mutation (false success).
Select a scenario above, then run the live test to walk the accountability pipeline.
UAR v2 proof center · action proof factors
// Intent + Policy + Tool + Transaction + Result → Action Proof (SHA-256) …
Open-source MCP-SEC-SCORE benchmark — reproducible via Docker, no login required
Run Repro BenchmarkNo Login · No API Keys · Instant Sandbox
One click runs a live FinanceBot scenario: authorized invoice lookup vs unauthorized database export — intercepted, revoked, and cryptographically verified.
$ financebot run --intent "Check August Invoice #8291"
Phase 3 · Real state & evidence
Four governed tool proposals: two legitimate reads, two blocked attacks. Each step seals a SHA-256 Universal Action Receipt via the same runtime path as production intercept.
SHA-256 UAR evidence (selected step)
Run the trajectory to mint receipts. Click a step to inspect its evidence bundle.
Live Dashboard Quick Nav
SEE → CONTROL → VERIFY → PROVE — intercept, govern, verify outcomes against systems of record, prove with UAR v2.
Threat intel, red team, trust hub, and compliance extend governance — they are not separate security products.
Secondary — Security Engines
Legacy guardrails remain available; primary product value is agent action governance, Universal Action Receipts, and verification APIs.
Live Interceptor Preview · Secondary Security Engine
Send adversarial prompts and PII payloads — watch Nexus Shield detect threats, block injection, mask secrets, and produce inspection evidence in real time.
50 of 50 free scans remaining
Routing: OpenAI GPT-4o
POST /api/sandboxSimulate an agent tool call and inspect a live SHA-256 Universal Action Receipt — before/after state hashes and tamper-evident SHA-256 receipt core (hash integrity only).
Receipt JSON will appear here after simulation.
Open-source harness benchmark metrics (Harness test suite (open-source baseline fixture — not live fleet telemetry)) — separate from your production UAR ledger on the data plane.
*Tested across 500+ MCP attack scenarios & multi-agent execution graphs. Open-source benchmark harness available on GitHub — OWASP GenAI / Agentic tags on every scenario.
measured in Nexus benchmark harness test environment
harness/fixtures/sample-output.json.Prove It — Live Attack → Block → Evidence
Agent Attack Attempt
Unauthorized shell command: rm -rf / — PII exfiltration via read_customer_records
Deployment simplicity, air-gapped guarantees, and CISO-ready signals — without shipping your prompts to third parties.
Full Trust PagePrompts, tool arguments, and action payloads are evaluated inside your VPC, cluster, or on-prem boundary. The data plane does not require Nexus Cloud, third-party LLM APIs, or outbound product telemetry for governance decisions and UAR sealing.
Nexus Shield never exfiltrates prompt or action payload data to Nexus-operated systems.
Every governed attempt follows the same isolated pipeline; Security Engines are optional supporting layers, not a substitute for action receipts.
┌────────────────── Customer infrastructure (air-gap capable) ──────────────────┐ │ │ │ Agent ──► Intent ──► Authority ──► Policy ──► Decision ──► Action ──► UAR │ │ (declare) (RBAC) (engine) (ALLOW/…) (execute) (SHA-256) │ │ │ │ ◄── No required egress to Nexus Cloud · optional local LLM only ──► │ └─────────────────────────────────────────────────────────────────────────────────┘
Deep dive: Architecture whitepaper · Data plane vs control plane
One helm upgrade --install deploys the Fast API governance plane, optional ML security engine, ingress, and health probes — suitable for EKS, AKS, GKE, and private clusters (GitOps / Operator-style lifecycle via Helm + ExternalSecrets).
Reference stack nexus-reference-app and docker-compose.nexus-reference.yml spin up the sidecar, mock MCP, and policy path in minutes for labs and light deployments.
API keys and sensitive config bind to Kubernetes Secrets, HashiCorp Vault (External Secrets Operator), or AWS Secrets Manager via the same ESO pattern — never committed to git.
Helm values (externalSecrets) →Nexus Shield does not leak prompt or action payload content to Nexus-operated services. On-prem and air-gapped modes keep inspection and UAR sealing on your metal.
CISO-ready architecture, control-plane separation, and benchmark vs production proof lanes — see the architecture whitepaper.
architecture-whitepaper.md →Governance runtime (MIT) and evaluation harness (Apache 2.0) are source-available for audit; container images support standard SBOM tooling (Syft, Trivy) in your registry pipeline.
harness repository →Try policy and sandbox flows without API keys: production landing playground (/api/sandbox) and local reference stack on :8090.
Open playground →Protect agents by count and intercepted tool calls — from sandbox engineers to enterprise FinTech control planes with cryptographic proof vaults.
Save ~17% with annual billing — Action Governance (Team) from $249/mo
Individual Engineers & Sandbox
1 Protected Agent | 10k Intercepted Actions
Discover agents, run hijack simulations, and inspect evidence bundles — no credit card.
B2B SaaS & AI Agent Platforms
Up to 25 Protected Agents | 500k Intercepted Actions
Production intent/action divergence checks, capability revocation, and SHA-256 UAR evidence for every governed action attempt.
FinTech, Insurance & Regulated Infra
Unlimited Agents | Custom Edge Sidecar
Inter-agent delegation governance, private edge deployment, and audit-grade evidence verification.
All plans include HTTPS, webhook signing, and org isolation. Prices in USD. Taxes may apply.
Kurumsal sorularınız, demo talepleri ve ortaklık fırsatları için ekibimize doğrudan ulaşabilirsiniz.
E-posta bağlantısına tıklayarak varsayılan posta uygulamanızı açabilir veya kopyala düğmesiyle adresi panoya alabilirsiniz.