Capability Mapping
SOC for AI · Platform Governance
← Back to Main
Kindo × Deloitte · SOC for AI

Platform Governance — Capability Mapping.

What the platform provides today and what we build next — prepared for Friday's working session with Kishore's team.

01 Context

Following Kishore's clarification of SOC for AI objectives, this document maps existing Kindo platform capabilities against each objective. Shadow AI discovery (finding unknown AI tools across the enterprise) is handled by Deloitte's detection engineering team through SecOps use cases. SOC for AI for Kindo focuses on monitoring and governing the agents built on the platform — detecting unauthorized changes, behavioral drift, and data contamination.

Delivery Velocity
This capability mapping was produced within 4 hours of receiving Kishore's requirements — leveraging existing platform documentation, API surface analysis, and governance architecture verification from prior sprints. Requirements mapping, dependency identification, and capability assessment run continuously.
02 Objective-by-Objective Mapping

Each of Kishore's five SOC for AI objectives mapped against current platform capabilities, identified gaps, and sprint targets.

Objective 1 Detect Unauthorized Agent Deployment / Updates Sprint 2
What Kindo Has Today
  • Agent inventory API (GET /v1/agents/list)
  • Audit logging for agent create / update / delete
  • RBAC (User / Agent Creator / Administrator)
  • Agent version history + restore (shipped)
  • Per-client RBAC isolation via Cerbos policies ("most restrictive wins")
What Needs Building

Alerting workflow — detection rules that fire when an unauthorized user creates or modifies an agent. The data exists; the monitoring layer does not.

Sprint Target

Sprint 2 — build the alerting and detection rules layer on top of existing audit data.

Partial
Odin's Assessment
Has: Agent inventory (GET /v1/agents/list), audit logging for workflow/agent create-update-delete, full version history/restore. Missing: Out-of-the-box alert for "unauthorized user created/changed agent" — data exists, detection workflow needs to be built on top.
Linear: ENG-8737 · Improved audit logging · Backlog / Medium
Objective 2 Detect Unauthorized Integration / Data Source Connections Sprint 2
What Kindo Has Today
  • Integration connection inventory API (GET /v1/integrations/connections)
  • Org-wide + per-user-group Tool Action Access Controls
  • Audit logging for tool server config + tool invocations
  • Pinned connection handling
What Needs Building

Scope monitoring — cross-reference agent tool calls against allowed integrations per SOP. Current controls prevent unauthorized access but don't detect scope violations within authorized integrations.

Sprint Target

Sprint 2 — scope violation detection against SOP-defined integration boundaries.

Partial
Odin's Assessment
Has: Integration connection inventory (GET /v1/integrations/connections), org/user-group tool access controls, pinned connection handling, audit logs for tool invocations. Missing: First-class alerting for "new sensitive integration added" or "agent started using new data source."
Linear: ENG-10042 · Org-level default settings + tool access controls · Backlog / High
Objective 3 Detect Behavioral Drift from SOPs Over Time Sprint 3
What Kindo Has Today
  • OTEL traces (workflow runs, step names, task durations, tool calls)
  • Hatchet workflow spans
  • Structured JSON logging with trace correlation
  • Successful / failed tool call audit logging
  • DLP filter events
  • LangSmith / Logfire export path (done, next build)
What Needs Building

SOP compliance monitoring engine — compares agent actions against SOP / skill file definitions over time, flags deviations. This is the primary build target and the most IK-dependent objective. No behavioral baseline or anomaly detection exists today.

Sprint Target

Sprint 3 (design in Sprint 2) — requires IK design session with SME to define SOP compliance baselines.

Weak Today
Odin's Assessment
Has: Logs successful/failed tool calls, LLM requests, sandbox code execution, DLP events. Missing: Behavioral-baseline / anomaly-detection / SOP-drift engine — this is the biggest gap. IK approach designed specifically for this gap.
No direct Linear ticket · IK approach designed for this gap
Objective 4 Detect Guardrail / Policy Changes Sprint 2
What Kindo Has Today
  • Audit log captures ALL administrative changes: security settings, tool-server config, user groups, DLP patterns, model management, organization settings
  • Full metadata with timestamp + acting user
What Needs Building

Change detection and alerting — automated diff and notification when guardrails or policies change. Data exists; monitoring layer does not.

Sprint Target

Sprint 2 — automated diff + notification on policy changes.

Partial
Odin's Assessment
Has: Audits admin changes — security settings, tool-server config, user groups, DLP patterns, model management, org settings. Missing: Built-in drift monitoring that automatically diffs and alerts on policy changes.
Linear: ENG-8539 · Principle of Least Privilege · Backlog  |  ENG-8654 · Security primitives · Backlog / Low
Objective 5 Detect Cross-Tenant Data Contamination Sprint 3
What Kindo Has Today
  • Tenant-scoped memory
  • Strict authorization around integration connections
  • Cross-tenant referent checks on workflow import
  • Pinned-connection logic (prevents silent credential substitution)
  • Per-client DLP controls
  • Per-client RBAC isolation on single instance
What Needs Building

Cross-org detection — "agent for client A referenced client B data" beyond audit logs + existing guardrails. Prevention is strong; detection is the gap.

Sprint Target

Sprint 3 — depends on multi-tenancy architecture confirmation.

Prevention > Detection
Odin's Assessment
Has: Tenant-scoped memory, strict authz around integration connections, cross-tenant referent checks on workflow import, pinned-connection logic. Missing: Dedicated detector for "agent for client A referenced client B data" beyond audit logs + existing guardrails. Stronger on prevention than detection.
Linear: ENG-8453 · Multi-Tenancy & Isolation · Adjacent
03 Infrastructure Already in Place

The building blocks for SOC for AI governance monitoring are largely deployed. The gap is the monitoring and alerting layer — not the underlying data.

Audit Log
Admin UI + CSV export + SMK/syslog forwarding to SIEM
OTEL Observability
Traces, metrics, and logs across all services
Grafana Dashboard Metrics
Chat messages, token usage, ingestion
Agent Inventory + Integration Connection APIs
Programmatic access to agent and integration state
Agent Version History + Restore
Full version tracking with rollback capability
RBAC + Tool Action Access Controls + DLP Filters
Layered access control and data loss prevention
Tenant Isolation Safeguards
Runtime isolation, import validation, memory path scoping — prevention layer across all tenant boundaries

Linear tickets tracking governance enhancements: ENG-8737 (improved audit logging), ENG-10042 (org-level tool access controls), ENG-8539 (principle of least privilege), ENG-8654 (agent security primitives), ENG-8453 (multi-tenancy & isolation).

04 Sprint Roadmap
Sprint 1 — Current
Ends Jul 13
Scope confirmation + foundation
  • Scope confirmation with Kishore
  • Capability mapping (this document)
  • Revised questionnaire
  • Audit log API validation
Deliverable: Complete capability mapping against all five objectives
Sprint 3 — Jul 28 – Aug 10
Jul 28 – Aug 10
Objectives 3, 5 — advanced detection
  • Behavioral drift via IK / koan curriculum (Obj. 3)
  • Cross-tenant detection (Obj. 5)
  • IK design session with SME
Deliverable: SOP compliance monitoring design + cross-tenant detection prototype
05 Platform API Endpoints for Governance Monitoring

Kindo's REST API and audit infrastructure provide the data substrate for all five SOC for AI objectives. These endpoints enable programmatic governance monitoring.

Endpoint
Returns
GET /v1/agents/list
Agent inventory — agentId, name, creatorName, createdAt, modelsInUse, recentRunIds, lastRunAtUtc
GET /v1/agents/{agentId}
Agent detail + inputs + version history
GET /v1/integrations/connections
All connections — id, display_name, integration_config_id, is_preferred, created_at
GET /v1/runs/{runId}
Run results and status
GET /v1/runs/{runId}/evals
Step-level evaluation verdicts
GET /v1/models
Enabled model catalog
Audit Log (Enterprise)
CSV export + SMK/syslog forwarding for SIEM integration
06 Kindo Engineering Roadmap Alignment

Active and planned Linear tickets that directly support SOC for AI governance capabilities.

07 Approach

The approach to SOC for AI demonstrates the value of continuous requirements mapping and rapid capability assessment. When scope clarification arrived, the team produced a complete capability mapping against all five objectives within hours — drawing on existing platform documentation, API surface analysis, governance architecture verified in prior sprints, and engineering assessment. This velocity comes from maintaining living knowledge of the platform architecture, not from starting fresh each sprint.