Sprint 2 Review · SOC for AI

Here's what we said we'd do. Here's what we did. And here's what we need from you to shape Sprint 3.

Agent Telemetry — Alpha Inside Kindo

This is an alpha release — architected to graduate directly toward production inside the Kindo platform — demonstrating the telemetry capabilities discussed at the Jul 10 session. It represents months of research and prototyping in a final shape to drive requirements finalization. Your input refines it.


Sprint 2 Commitments & Delivery

Sprint 2 responded to the five governance objectives from Kishore's Jul 7 email and the direction from the Jul 10 co-design session. Every commitment addressed.

What We Said We'd Do

  • Build agent telemetry responding to Kishore's 5 governance objectives
  • Research the platform compatibility matrix Krishna asked for
  • Define the three-pillar discovery architecture with Kush and Charlie
  • Document the gateway approach Kush endorsed
  • Propose OTel/MLflow architecture for telemetry infrastructure
  • Begin discovery menu of options (Krishna's CrowdStrike analogy)
  • Scope digital twin, triage agent, and A7 quality audit for co-design

What We Delivered

  • Alpha release — trace export, LLM judges, eval config, span detail, error diagnostics
  • 5-platform matrix complete (Anthropic, Azure, Copilot, ServiceNow, Bedrock)
  • Three-pillar framework defined — Platform, Gateway, Network
  • Gateway architecture documented, Kush endorsed Jul 16
  • Two-tier OTel model proposed, alpha validated in dev — awaiting co-design
  • Discovery menu synthesis — awaiting co-design
  • Scoped and ready — awaiting co-design sessions to proceed

Kishore's Five Governance Objectives

Each objective mapped to a platform capability. Badges indicate honest status — what's built, what's research, what needs your input.

Objective 1

Detect Unauthorized Agent Deployment

Detect when unauthorized users are deploying new agents or updating existing agents.

Platform response: Pillar 1 — Platform Discovery (research complete) + Agent Telemetry (alpha). Kindo probes Anthropic, Azure, Copilot, ServiceNow, and Bedrock admin APIs to enumerate agents deployed across the enterprise — including those outside Kindo. Unauthorized agents are discovered, flagged, and routed into Kindo for governance. Once sanctioned and flowing through Kindo, Agent Telemetry provides continuous visibility: every agent in the catalog with creator, last updated, type, and telemetry export capturing every run.

We'd welcome your input on which platforms to prioritize for discovery integration.

Objective 2

Detect Unauthorized Tool/Data Connections

Detect when new integrations to unauthorized tools/sensitive data sources are added.

Platform response: Pillar 1 + Pillar 3 — Platform Discovery + Network Monitoring (research complete). Shadow AI connections — direct Anthropic API usage, Copilot integrations, unsanctioned endpoints — are discovered through platform API probing and network-level detection. Kindo orchestrates existing endpoint detection systems to surface unauthorized tool and data connections across the enterprise. For sanctioned AI flowing through Kindo, the Federated MCP Gateway enforces real-time tool access policies, and trace detail captures every tool call with arguments and data returned.

We'd welcome your guidance on defining "unauthorized" within your governance framework.

Objective 3 · Biggest Engineering Need

Detect Agent Behavioral Drift

Detect when actions performed are NOT in-line with the standard operating procedures defined for the agents.

Platform response: LLM Judges (alpha) — plain-language evals that score every agent run. Once workloads flow through Kindo — whether native or rerouted from discovery — judges flag behavioral deviations run-by-run against defined standard operating procedures.

To maximize this capability, we'd like to co-define which standard operating procedures and behavioral criteria the judges should evaluate against.

Objective 4

Detect Guardrail/Policy Changes

Detect changes in MCP/tool policies to provision edit/delete privileges to agents when not intended.

Platform response: Configuration panel — telemetry export, sampling rate, retention, and eval configs visible per agent.

Full change tracking and audit trail for configuration modifications is on the roadmap as a next step.

Objective 5

Detect Cross-Tenant Data Contamination

In the multi-tenant environment, detect data breach/contamination when agent for one client cross-references data of another client.

Platform response: Organization-isolated workspaces — traces stored per org, cross-tenant isolation enforced at the platform level.

Per-org isolation demonstrated in alpha. Production-grade cross-tenant monitoring is a planned enhancement.


Co-Design Input Requested

For the next session, we'd like to walk through these capabilities together and get your input on shaping what comes next.

We delivered what we could advance autonomously. What comes next, we build together.


Where Things Stand

Sprint 2 (Jul 13–25). We focused on what we could advance autonomously. Items that need collaborative input are waiting for the next co-design session — we're aligned with your pace.

Delivered Awaiting co-design input
Alpha

Agent Telemetry

Alpha release: trace export, LLM judges, eval configuration, span detail, error diagnostics. Architected to graduate toward production — deployment is a next step together.

Research Complete

Platform Compatibility Matrix

5-platform research (Anthropic, Azure, Copilot, ServiceNow, Bedrock). Based on public API documentation. Ready for your input on platform priority.

Research Complete

Pillar 2 Gateway Architecture

Kush endorsed gateway approach (Jul 16). Three-standard strategy documented. Architecture ready for collaborative refinement.

Awaiting Co-Design

OTel/MLflow Architecture

Two-tier model proposed: export (current alpha) vs native pillar. Alpha validated in development. Next steps depend on collaborative input.

Awaiting Co-Design

Discovery Menu of Options

Krishna-facing deliverable. 5–7 discovery methods with tier mapping. Ready for collaborative refinement.

Awaiting Co-Design

Digital Twin L2/L3 Analyst Pilot

Teams integration needs Deloitte IT coordination. Ready to pick up together in Sprint 3.

Awaiting Co-Design

Triage Agent Optimization

Requires work session with Zun. Ready to schedule when timing works for your team.

Awaiting Co-Design

A7 Quality Audit Agent Design

Depends on digital twin approach + design session with Krishna's team. Ready to co-design in Sprint 3.


Three-Pillar Discovery Framework

Developed collaboratively at the Jul 10 co-design session. Three approaches to AI governance — each platform has different strengths.

Pillar 1

Platform Integrations

Reach into Anthropic, Azure, Copilot, ServiceNow, Bedrock via admin APIs. Enumerate agents, audit usage, inspect configs.

Pillar 2

LLM Gateway

Kindo as inference proxy — 2 lines of config to redirect. Real-time governance at the choke point.

Pillar 3

Network Monitoring

Sweep for shadow AI — the long tail. Network and endpoint-level discovery beyond managed platforms.


Discovery Tier Model

Inspired by Krishna's CrowdStrike analogy — a menu of options, not one-size-fits-all. Proposed framework for co-design validation.

Basic

Detect & Respond Post-Action

Telemetry export + audit log analysis. After-the-fact detection. Low adoption cost.

Standard

Monitor & Alert Live

Native telemetry + LLM judges scoring every run. Real-time drift detection.

Elite

Preventative Blocking

Gateway position = real-time control. Block unauthorized actions before they execute.

Proposed framework, not yet validated with Deloitte.


Platform Compatibility Matrix

Per-platform capabilities for discovery and governance. Based on public API documentation — the research Krishna asked for.

CapabilityAnthropicAzure AI FoundryMS CopilotServiceNowAWS Bedrock
List models/agents✓ API✓ ARM API~ Graph catalog API ²~ Table API ³✓ API
Usage/audit logs✓ Usage/Cost + Compliance APIs ¹✓ Azure Monitor *✓ Purview✓ System logs✓ CloudTrail *
Content/safety filters✗ Not via API✓ Full API + Terraform~ DLP policies~ Admin config✓ Guardrails API
Disable/block access✓ Key + workspace revocation✓ RBAC + Policy✓ Licensing + Integrated Apps (+ Entra CA)✓ Role-based✓ IAM + SCPs
Enforce policies~ Spend/rate limits (no content policy)✓ Azure Policy✓ DLP + labels~ Workflows✓ Guardrails + SCPs
Alerting~ Spend notifications + SIEM polling✓ Monitor Alerts✓ Sentinel✓ Event Mgmt ⁴✓ EventBridge
Best Kindo methodAPI + Gateway (P1+2)API + Gateway (P1+2)API + Network (P1+3)API (P1) ⁵API + Gateway (P1+2)

Full API · ~ Partial · Not available

Notes
¹ Anthropic Compliance & Analytics APIs = Enterprise plan only; aggregated data, no prompt content.
² Copilot Graph catalog API in Public Preview (May 2026), requires Agent 365 license.
³ ServiceNow AI Agent tables require Pro Plus or Enterprise Plus SKU. Some tables have short retention (~13 days).
⁴ ServiceNow Event Management requires separate ITOM license.
⁵ ServiceNow LLM calls route from SN cloud, not customer network — P3 network monitoring effectively blind. P1-only.
* Azure Monitor and CloudTrail log metadata by default. Prompt/response content requires opt-in.
License Gates
The discovery "menu of options" is also a menu of license prerequisites: Agent 365 (Copilot), Purview premium (retention), Pro Plus/Enterprise Plus (SN AI Agents), ITOM (Event Mgmt), Anthropic Enterprise (Compliance/Analytics). Not every method works without the right licenses.

Three Open Standards

Kindo reaches governance outcomes on existing open standards — no proprietary protocol required.

Live Today

Inference APIs

Chat Completions, Responses, Messages. 2 lines of config to route any agent through Kindo.

Live Today

Federated MCP Gateway

Single governed endpoint for tool/integration access. Real-time control over agent tool use.

Landing Now

OpenTelemetry

OTLP-based. Kindo ingests traces from any AI workload. Foundation for evals and governance.


How We Got Here

Sprint 2 builds directly on Sprint 1 and the Jul 10 co-design session. Expand each card for full detail.

Sprint 1 — SOC for AI Proof of Concept

• Governance monitor agent scanned 286 agents, 45 integrations in <60 seconds • Covered all five governance objectives as a standalone PoC • Validated the concept — Deloitte asked for more depth

The team built a SOC for AI Governance Monitor as a triggered agent on Kindo. The agent scanned the full Kindo deployment, evaluated active integrations, and assessed change management controls. The result was a proof of concept demonstrating that AI governance monitoring could work at speed — setting the stage for the Jul 10 co-design session with Deloitte.

Jul 10 — Co-Design Session with Deloitte

• Kush: "You guys are phenomenal" — validated velocity, asked for gateway approach • Krishna: CrowdStrike analogy — "menu of options," tier model, platform matrix request • Kush: "Don't build a data lake" — architecture constraint that shaped Sprint 2

Presented to Kush, Krishna, Nathan, and Zun. The session defined every deliverable in Sprint 2:

Kush
"You guys are phenomenal. Really appreciate you guys turning this around this fast."
Tony's response: "We definitely don't want to make you look bad. You guys are 550,000 people and you have a structure to accommodate for that."
Kush — Gateway Philosophy
"If LLM is the choke point, what can we put there? Anything that runs within your four walls, if it has to go to your sanctioned LLM, that's the choke point."
Origin of Pillar 2 — Kindo as the LLM gateway.
Krishna — CrowdStrike Analogy
"We should have a menu of options — five, ten different ways. Each option has a pro and con."
Origin of the Discovery Menu and tier model.
Krishna — Platform Matrix Request
"Is there a list of platforms you were able to discover agents?"
Direct origin of the Platform Compatibility Matrix research.
Kush — Architecture Constraint
"You guys should not go down the path of building a data lake."
Noted. All telemetry uses existing Postgres + S3 — no new data lake.

Jul 7 — Kishore Defined the Scope

• SOC for AI = monitoring known AI agents, not shadow AI discovery • Five governance objectives defined (the foundation of this review) • Two requirements: audit logs/telemetry + data collection architecture

Kishore's email clarified a key misunderstanding: SOC for AI is about monitoring known AI agents and platforms — not shadow AI discovery, which is handled by Deloitte's detection engineering team. He defined the five governance objectives that became the foundation of Sprint 2 and specified two implementation requirements: (1) audit logs and telemetry data from the platform, and (2) a process and data collection architecture to monitor for these risks.

Sprint 2 — What We Set Out to Do

• Platform matrix Krishna asked for • Three-pillar architecture defined with Kush and Charlie • Telemetry capability that makes all of it possible

Take the feedback from Jul 10 and Kishore's objectives, and respond with working demonstrations: the platform compatibility matrix that Krishna requested, the three-pillar discovery architecture defined collaboratively by Charlie and Kush, and the agent telemetry capability — built natively inside Kindo — that makes governance monitoring concrete and testable.