Five layers of Zero Trust.
Built for AI agents.

Every request is authenticated, authorized, content-scanned, and immutably logged before any data is returned. No exceptions.

Ed25519 PKI · mTLS 1.3 · HSM

Self-contained CA issues ephemeral 1-hour leaf certs at first connection. TLSv1.3 minimum on every hop. PKCS#11 HSM keeps root keys off the process heap. Replay attacks rejected via Redis nonce window (300 s default).

  • Ed25519 PKI
  • mTLS 1.3
  • HSM / PKCS#11
  • Replay Protection
  • AES-256-GCM at Rest
cert_ttl 1 h
tls_version TLSv1.3
key_algo Ed25519
replay_window 300 s
at_rest AES-256-GCM
agent_cache
LRU 10 000 slots TTL 60 s
posture_score
1.0 ✓ active 0.4 ✗ low → 403 quarantined → 403
agent_fsm
active → quarantined → decommissioned

Zero Trust Posture Engine

Every incoming request carries a posture score derived from MITRE tactic risk. Quarantined or decommissioned agents are rejected with 403 before any policy evaluation. An LRU cache (10 000 slots, 60 s TTL) prevents unbounded memory growth under high agent traffic.

MITRE ATT&CK + ATLAS Dual Framework

Every tool call is tagged with ATT&CK Enterprise and ATLAS tactic IDs. A Redis Stream records call rate, payload size, and tool diversity per agent over a rolling 7-day window. Statistical deviation from the established baseline triggers a graduated response: warn at 50%, auto-quarantine at 75%. High-risk tactics (Exfiltration AML.TA0009, Impact AML.TA0010, Execution AML.TA0004) receive an immediate risk boost.

mitre_frameworks
ATT&CK Enterprise ATLAS AML.*
baseline_window
7 days · Redis Stream
thresholds
warn ≥ 50% quarantine ≥ 75%
high_risk_tactics
AML.TA0009 · AML.TA0010 · AML.TA0006 · AML.TA0004
input_scans
SQL injection Prompt injection Path traversal SSRF / JNDI CMD injection
output_redaction
Email Phone IBAN Credit card CIN
payload_limit
512 KiB default · configurable

AI Shield — Prompt & PII Guard

Every tool call is scanned for prompt injection, SQL injection, path traversal, SSRF/JNDI, and command injection. Outputs are redacted for PII: email, phone, IBAN, credit card, and national ID. Optional Presidio NLP enhancement runs in a thread pool to avoid blocking the event loop. Payload size is hard-blocked at 512 KiB by default.

Immutable SHA-256 Hash Chain

Every audit event is appended to a tamper-evident SHA-256 chain. Each record includes prev_hash, a deterministic content hash, nanosecond timestamp, actor, and OTel trace ID. ChainVerifier detects any tampering. JSONL export for SIEM. Compatible with WORM storage. No DELETE ever.

hash_algo
SHA-256
chain_fields
prev_hash · content_hash · nanosecond_ts · actor · trace_id
export
JSONL OTLP SIEM WORM
mutation_policy
NO DELETE · NO UPDATE

OPA Policy Engine — per tool, per tenant

Every tool call is evaluated against Open Policy Agent before execution. Gate by tenant, tool name, product scope, and time-of-day. OPA bundle hot-reloaded without gateway restart. Fail-closed: if OPA is unreachable, all calls are denied by default.

tenant isolation tool allowlist product scope time-based policies fail-closed default hot-reload no restart

Security Fabric

v0.3 extends the five Zero Trust layers with an operational security fabric: a unified Security Center, agentic circuit breaking, scoped API keys, source scanning, shadow policy governance, and Merkle-sealed audit evidence.

Security Overview / Security Center

A single pane of glass for gateway security posture: a global security score, critical alerts, expired or revoked keys, authentication failures, and OPA policy violations — aggregated in real time.

global security score critical alerts expired / revoked keys auth failures OPA violations

Agentic Circuit Breaker

Per-session budgets stop runaway agents before they burn resources: call count, repeated calls to the same tool with identical arguments, session duration, and token consumption. Operators can trigger a manual emergency stop from the dashboard.

call count same tool+args repetition duration tokens manual kill switch

Scoped API Keys

Keys are created with a mandatory expiration (max 90-day TTL), support rotation and immediate revocation, and are bound to read / write / admin / MCP scopes. Only hashes are stored, and the plaintext key is displayed exactly once at creation.

TTL ≤ 90 days rotation instant revocation read / write / admin / MCP scopes hash-only storage

Source Security Scanner

Every registered source is vetted before it can serve tools: TLS verification, certificate expiry checks, SSRF blocking of private IP ranges, and a scheme whitelist. Secrets are referenced from Kubernetes or Vault — never stored in plaintext.

TLS verification cert expiry SSRF private-IP block scheme whitelist K8s / Vault secret refs

Shadow Governance

Evaluate new OPA policies in shadow mode — without enforcement — before rolling them out. The impact report shows exactly which calls would flip from allow to deny and from deny to allow, so policy changes ship with zero surprises.

shadow OPA evaluation no enforcement allow→deny / deny→allow report

Merkle-Sealed Audit Vault

The audit chain is periodically sealed with Merkle roots, on top of the per-record SHA-256 hash chain. Auditors can verify integrity offline against the published seals — cryptographic proof that no event was altered, reordered, or deleted.

hash chain Merkle seals offline verification

Behavioral Tracking & Auto-Quarantine

Each agent builds a behavioral baseline over time — call rate, payload size, tool diversity. When live behavior deviates 75% or more from that baseline, the agent is automatically quarantined and cut off with 403 responses.

per-agent baseline warn ≥ 50% auto-quarantine ≥ 75%

Security at a glance

ArcaX security is enforced on every single tool call, at every layer, with no bypass possible for authenticated tenants. Transport: Ed25519 PKI, mTLS 1.3, 1-hour ephemeral certs, HSM PKCS#11 key storage, replay protection via Redis nonce window. Identity: Zero Trust posture score per agent, LRU-capped cache, agent FSM (active → quarantine → decommission). Behavior: MITRE ATT&CK Enterprise + ATLAS dual-framework tagging, Redis Stream 7-day behavioral baseline, statistical deviation scoring, auto-quarantine at 75%. Content: AI Shield scans every input for injection patterns (prompt, SQL, path traversal, SSRF) and redacts PII from every output. Forensics: Immutable SHA-256 hash chain, no DELETE ever, WORM-compatible JSONL export, OTel SIEM delivery. Policy: OPA evaluated per call, per tenant, per tool, fail-closed.

View Compliance Read Security Docs

Ready to connect your agents to your enterprise?

Get a personalised demo and a project-specific ROI estimate. No commitment. No cloud dependency required.

Request a Demo