EXPOSURE // INCIDENT ZERO

Governed AI incident response where ChatGPT investigates infrastructure through WebMCP, humans authorize consequential actions, and the agent executes remediation and verifies recovery.

Comment

Inspiration

Production incidents rarely fail in obvious ways. A database may appear overloaded even when the real cause is an application deployment several layers upstream. Engineers must correlate telemetry, logs, dependencies, changes, risk, and human constraints—often while customers are already affected.

Exposure explores a better operating model: an infrastructure engineer and an AI agent working together inside the same live command environment.

What it does

Exposure is an agent-native operational intelligence system built around a simulated nine-node Enterprise Linux 9 production fleet.

The featured scenario, Incident Zero, begins when a checkout platform develops severe latency and HTTP 502 errors. PostgreSQL appears to be failing, but the database is actually a downstream victim of a faulty application deployment.

Using WebMCP, the agent can:

  • Inspect fleet health
  • Examine individual hosts and services
  • Map infrastructure dependencies
  • Query metrics and logs
  • Review deployments and configuration changes
  • Develop and test competing root-cause hypotheses
  • Record human operational constraints
  • Evaluate remediation options
  • Request explicit human authorization
  • Execute a simulated rolling rollback
  • Verify recovery through measured evidence
  • Generate an auditable incident record

Why WebMCP

WebMCP is the operating interface between Exposure and the AI agent. Instead of requiring ChatGPT to infer actions from page text or manipulate arbitrary UI controls, Exposure exposes explicit, governed incident-response tools with defined inputs, state transitions, safety constraints, and authorization boundaries.

The agent can investigate the incident, register constraints, propose remediation, request authorization, execute an approved rolling rollback, and verify recovery. It deliberately cannot authorize its own consequential remediation.

This creates a practical human-agent operating model: the AI can reason and act, while the human retains authority over consequential decisions.

Human authority

Exposure does not allow an agent to jump directly from suspicion to execution.

The operator can establish a constraint such as:

Do not restart PostgreSQL.

That constraint becomes part of the operational state and blocks conflicting actions. When Exposure identifies the faulty checkout deployment, it proposes a rolling rollback and stops at an authorization gate.

Only the human operator can approve the consequential action.

Incident Zero

The simulation models:

  • One Nginx edge load balancer
  • Two Nginx web servers
  • Three checkout application servers
  • One Redis cache server
  • One PostgreSQL primary
  • One PostgreSQL replica

A faulty checkout release creates excessive database connections across the application tier. This causes database pressure, replication lag, web-tier timeouts, and customer-facing failures.

The agent must distinguish root cause from downstream symptoms, compare remediation options, respect the operator’s constraints, and prove that recovery occurred.

Product experience

Exposure presents the incident through a cinematic operational interface:

  • A living infrastructure topology
  • Animated traffic and dependency paths
  • Fleet telemetry
  • Structured evidence
  • Competing hypotheses
  • Contradiction detection
  • Risk assessment
  • Human authorization gates
  • Visible rollback progression
  • Recovery verification
  • A sealed decision trace

Every visual transition communicates real application state rather than decorative animation.

Technical approach

Exposure uses TypeScript, React, Vite, structured JSON Schema inputs, and the WebMCP document.modelContext.registerTool() API.

A deterministic client-side incident engine controls fleet state, evidence, tool results, authorization, remediation, and recovery. The competition build requires no production credentials, external infrastructure, or paid APIs.

The fleet is simulated, but its telemetry, Linux services, logs, dependencies, failure behavior, and remediation sequence are designed to remain technically credible.

What comes next

The challenge release focuses on one polished, deterministic incident. Future versions could connect Exposure to authenticated operational systems such as Prometheus, Grafana, Ansible, cloud platforms, service-management systems, and enterprise observability tools.

Exposure’s objective is simple: expose what is happening, reveal why it is happening, control what happens next, and prove whether the action worked.

Built With

Share this project:

Updates

Submission history