Public documentation for governed AI labor
SDKs/Governance/Connectors
Arx / Docs / kyle-platform-update.md

Documentation

kyle-platform-update.md

Project-Agent-trust-merge / outreach/kyle-platform-update.md

Project-Agent-trust-merge market outreach/kyle-platform-update.md

Subject: What security AI agents get when they deploy on ARX — Wiz, Endor, Saviynt, Snowflake, API & ePDF

Hey Kyle,

Quick update on what ARX actually gives a security AI agent the moment it's deployed on the platform, plus the integrations I know you've been tracking: Wiz, Endor Labs, Saviynt (identity governance), Snowflake (data governance), and our API and ePDF layers.

What a security AI agent gets on Day 1

The instant an agent is wired into ARX, it inherits the full control plane without writing any of it itself. Concretely, the agent gets:

  • Credential-free tool access. It never touches API keys. ARX injects credentials from the encrypted vault at execution time against 101 pre-built connectors and roughly 2,500 governed operations.
  • A policy-evaluated execution path. Every call passes through the intercept → evaluate → log pipeline. Each operation is risk-scored and returned as PERMIT, ESCALATE, or DENY — so the agent can't quietly escalate blast radius.
  • Human-in-the-loop by default. Escalations route to Slack or Teams with one-click approve/deny. The agent waits; the audit captures the decision.
  • An immutable audit trail. Every action, verdict, risk score, and result is written append-only and streams to Splunk, Sentinel, syslog, or webhook.
  • Behavioral drift detection. The agent declares an intent manifest; anything outside it gets flagged, escalated, or denied.
  • Compliance artifacts for free. SOC 2, ISO 27001, and FedRAMP evidence is derived from actual platform data, not hand-authored.

Deployment process

Deployment is intentionally a ~10-minute path so teams can stand an agent up without a services engagement:

  1. pip install agentvault-sdk (or install the JS SDK).
  2. Create an API key in the ARX console under Settings → API Keys with the deployer role.
  3. Export ARX_API_KEY and ARXSEC_API_URL and instantiate ARXClient() — the SDK picks them up automatically.
  4. Call the typed connector factory (e.g. arx.crowdstrike(), arx.wiz(), arx.snowflake()) and execute an operation. Under the hood this hits POST /v1/sdk/execute, which resolves credentials, evaluates policy, logs the action, and executes against the target.
  5. Verify in the audit trail, then create a policy rule (UI or POST /v1/policies/rules) to lock down any high-risk operation behind an ESCALATE verdict.

That's it — the agent is now governed end-to-end. No custom audit pipeline, no secrets plumbing, no approval service to build.

Wiz

A Wiz-connected agent can pull CRITICAL/HIGH issues, triage vulnerabilities, walk cloud resource graphs (search and relationship traversal), read security controls and frameworks, pull reports, and update or resolve issues — all OAuth2, token-cached, and policy-gated. The showcase workflow, Wiz Finding Distribution, pulls findings, routes them to the right team channel based on ownership, and opens Jira tickets automatically. Cloud Misconfiguration Remediation sits on top of the same connector and closes the loop on common fixes.

Endor Labs

An Endor-connected agent gets reachability-based supply chain triage: findings with reachability and call-path data, projects, packages with full dependency trees, policies, scans, CI/CD rules, repositories, and metrics. The payoff is that agents can actually prioritize — Endor Finding Triage drops unreachable vulns, tickets the rest, and pages on anything critical + reachable. We also ship SBOM export, dependency-risk alerting, SCA policy-violation response, and container-image scan gating as prebuilt workflows.

Identity governance — Saviynt

Saviynt is the IGA layer. A Saviynt-connected agent can read identities, accounts, and entitlements; run and respond to certification campaigns; request, approve, and revoke access; create and update roles and policies; and query audit events. Destructive actions — disabling accounts, revoking roles, bulk decertifications — are risk-scored and ESCALATE-by-default so a human signs off. The flagship workflow, Access Certification Campaign, pulls stale access and excessive group memberships from Saviynt, routes reviews into ServiceNow, and collapses a two-day quarterly exercise into roughly two hours. Stale Account Deactivation and MFA Enforcement Check ship alongside it. ARX itself also enforces RBAC (Admin / Deployer / Auditor / Viewer) with SCIM 2.0 group-to-role mapping.

Data governance — Snowflake

Snowflake is the data governance layer. A Snowflake-connected agent can inspect object tags and classification, read and manage masking and row-access policies, walk access history and object dependencies for lineage, and query the governance views over warehouses, databases, schemas, and shares. Agents can discover sensitive columns, propose or apply masking and row-access policies, and pull audit evidence straight from access history — every mutation is ESCALATE-by-default. The Compliance Evidence Collection workflow assembles classified objects, policy assignments, and access history into the audit package, saving roughly eight hours per audit.

API integration

Everything the SDK does is available over the REST API, which is how most agents, CI/CD pipelines, custom dashboards, and webhook handlers integrate. Agents get:

  • POST /v1/sdk/execute — the single entry point for every governed operation, with session context and auto-populated prior actions.
  • /v1/audit — filter, page, and stream the audit trail.
  • /v1/policies/rules — create and manage policy rules programmatically.
  • /v1/compliance — list, fetch, and export compliance packages.
  • /v1/api-keys — mint scoped, expiring API keys for each agent.
  • Bearer JWT or X-API-Key auth, per-org rate limits with X-RateLimit-* headers, consistent JSON envelope (data + meta.request_id on every response), and standard status-code semantics — including the policy-verdict mapping (PERMIT → 200, ESCALATE → 202/403, DENY → 403).

Net effect: agents in any language, or any system that speaks HTTP, can be governed by ARX without waiting on an SDK.

ePDF integration

ARX generates branded, SOC 2-ready PDF evidence packages directly from platform data — no manual authoring. A single call to GET /v1/compliance/{package_id}/pdf returns a PDF containing the agent's execution graph, data flow diagram, VSQ responses, sub-processor list, and SOC 2 control mappings, with the agent name, generation timestamp, and requesting user stamped in the header. Audit logs export the same way: GET /v1/audit/export?format=csv|json produces auditor-ready evidence with verdicts, durations, and resolved user names. For agents, this means the compliance artifact auditors actually ask for is a one-API-call away, not a quarterly scramble.

Wrapping up

Deploying on ARX turns a raw security AI agent into a governed one: credentials handled, policies enforced, humans in the loop for the scary stuff, every action auditable, and compliance evidence generated on demand — with Wiz, Endor, Saviynt, and Snowflake covering the cloud security, supply chain, identity, and data surfaces out of the box.

Happy to walk through the 10-minute deployment path live, or show any of the Wiz / Endor / Saviynt / Snowflake workflows end-to-end.

Thanks,