SOC for AI · Kindo × Deloitte · Monday, August 17

Strategic Frame

Transformation Platform, Not Point Solution

Kindo is not another tool to help humans do security a little faster. It is a transformation platform.

Inside → Out

Kindo works from inside the customer's environment outward. On-prem, air-gapped, local models, local data. The customer's infrastructure is the foundation — Kindo builds on top of what they already control.

Outside → In

Palo Alto, CrowdStrike, Microsoft work from outside the perimeter inward. Network intercept, cloud scanning, endpoint monitoring. The vendor's infrastructure is the foundation — the customer routes through them.

Deloitte's Strategic Autonomy

With Kindo, Deloitte can run whatever partner fits best, whatever makes the most money, whatever the customer wants — and Deloitte stays in control. They don't turn it over to Palo Alto or anyone else. Kindo is the control plane that Deloitte operates — partners plug in below.

Competitive Position

Hard Moat · Soft Moat

Hard moat = competitors don't have it. Soft moat = switching costs protect even with partial overlap. Red = concede to the big players.

Hard Moat — Unique to Kindo

Differentiator #1

On-Prem / Air-Gapped

Full K8s Helm charts, no internet required. Local models, local compute, local data. AIRS has partial network intercept only.

● Live

Unique Capability

Agent Execution Engine

Action Bots, Action Chat, triggers, code gen, RAG, memory, human-in-the-loop. No competitor builds and runs agents.

● Live

Security + Economics

Turbo Mode

Compile validated agent patterns into deterministic code. No hallucination, no prompt injection surface, no model drift. Cost → near-zero.

● Live

Isolated Compute

VMI Sandbox

SQL, Object Store, Vector DB, local files, DeepHat model — fully isolated execution environment per agent.

● Live

Partnership Lock-In

Deloitte-Specific Customization (CDA)

Features built to Deloitte specs, CDA-native. Forward Deployed Engineering model — T&C owns, maintains, and supports everything built on-prem.

● Live

Soft Moat — Advantage with Switching Costs

Live

MCP Gateway

Federated MCP with RBAC, 200+ tools. Agent context makes tool selection smarter than competitors.

● Live

Live

Inference Proxy

LLM proxy, DLP, audit logging, 26+ models. Model-agnostic vs vendor lock-in.

● Live

Live

Governance Layer

Policy, DLP, audit logs, Secrets Mgr, Admin Portal. Table stakes across vendors.

● Live

Live

Session Context

Full session history. Cross-session patterns. CrowdStrike has this within Falcon.

● Live

Soft Moat In Development — Krishna’s Sales Motion

The Demo

Anomaly Detection

Baseline per agent, drift detection. The execution layer catches rogue behavior a gateway can't see.

Oct

Aug Build

Audit Trail → SIEM

OTel export to Google SecOps, Splunk, Datadog. Every agent action in the client's audit log.

Aug

Sep Build

Governance Dashboard

Active agents, cost, tools in use, compliance. Single-pane operator view.

Sep

Sep Build

FinOps / Cost Controls

Token tracking, budgets, cost attribution. Model-agnostic vs M365-only.

Sep

Concede — Not Our Arena

Concede

Shadow AI Discovery

Network-layer discovery of unsanctioned AI. Requires DPI — Kindo is app-layer.

Do Not Build

Concede

AI Model Scanning

Model vulnerability scanning, supply chain analysis. Requires massive ML security investment.

Do Not Build

Concede

Automated Red Teaming

Adversarial testing of AI models and agents at scale. Prisma AIRS and Microsoft invest heavily here — not viable at Kindo’s scale.

Do Not Build

Features

6 Features — What CDA Needs

Each feature answers a specific CISO question. 2 live, 4 to build. Mapped to the four-pillar strategy.

01

AI Inference Gateway

All CDA agent traffic routes through Kindo. Model access control, DLP, audit logging.

● LiveP1

"Who controls which models our agents use?"

02

Tool Access Control (RBAC)

MCP gateway with Cerbos RBAC. Which agents get which tools is policy, not code.

● LiveP2

"Can an agent access systems it shouldn't?"

03

Client-Configurable Policies

Per-client policy rules. "Analysts can query prod but not export." Configurable by operators.

AugP4a · Critical Path

"Can we set our own governance rules?"

04

Audit Trail to Client SIEM

OTel telemetry export to Google SecOps, Splunk, Datadog. GenAI semantic conventions.

AugP3

"Can we audit every AI decision in our SIEM?"

05

Governance Dashboard

Single-pane: active agents, cost, tools in use, policy compliance status.

SepP1+P2+P3

"Show me what our AI agents are doing right now."

06

Behavioral Anomaly Detection

Agent A normally calls 3 tools. It just called 47. Only the execution layer catches this.

OctP4b · The Demo

"What if an agent goes rogue?"

Roadmap

What Krishna Can Sell

Quarterly delivery mapped to CDA capabilities. Each milestone adds sellable features.

Now

Sellable Today

✦ AI Inference Gateway

✦ Tool Access Control (RBAC)

✦ On-prem / air-gapped

✦ Agent execution

✦ Turbo Mode

Why it matters: CISO can answer “who controls our AI?” today. Air-gapped = no data leaves the building. Turbo Mode = validated agents become deterministic code — zero hallucination risk.

Sep

Sprint 1

→ Client-configurable policies

→ Audit trail to SIEM

Why it matters: Clients set their own governance rules without code changes. Every AI decision lands in their existing SIEM (Google SecOps, Splunk, Datadog) — audit-ready from day one.

Oct

Sprint 2

→ Governance dashboard

→ FinOps / cost controls

Why it matters: Single-pane view across all agents, tools, models, and cost. Model-agnostic — unlike Microsoft’s Agent 365 which only covers M365 workloads. Budget controls prevent runaway LLM spend.

Nov

Sprint 3

→ Behavioral anomaly detection

→ Full SOC for AI live

Why it matters: The demo. Agent A normally calls 3 tools — it just called 47. Only the execution layer catches this because it knows the baseline. A gateway can’t. This is the rogue-agent story that sells.

Market Momentum

The Demand Signal Is Here

The market shifted in 12 months. Enterprises are moving from “what is this?” to “who do I buy from?”

Aug ’25

Black Hat — Last Year

Nobody at the booth. SOC for AI was a concept, not a budget line.

Aug ’26

Black Hat — This Year

Lines 4–5 people deep at all four demo stations. 500+ deep conversations.

In six months, enterprises will have a plan and be ready to buy. The window to establish CDA as the answer is now.

Objectives

Kishor's 5 SOC for AI Objectives

Each objective maps to Kindo capabilities — native and partner.

1

Detect new agent deployment

● Live

2

Detect new tool / data source connections

● Live

3

Detect behavioral drift in agent actions

Oct

4

Detect guardrail / policy changes

Oct

5

Multi-tenant data contamination

Scoping