AegisFleet: Autonomous Google Cloud SOC Tier 1 Responder

• 

  Automated Incident Triage & AI-Generated Narratives

• 

  Mapping the Attack Path with Dynamic Visualizations

• 

  Human-in-the-Loop: Deterministic Staged Containment

• 

  Failsafe Operations: The 1-Click Rollback Engine

• 

  Integrating with Workflow: Bidirectional ChatOps & Slack Block Kit

• 

  Automated Executive Reporting: Instant CISO Briefings

• 

  Transparent AI: Real-Time Agent Swarm Execution Trace

• 

  Multi-Stage Remediation: Neutralizing Compromised Credentials

• 

  State-Aware Safety: Intelligent Rollback Tracking

• 

  Translating Technical Threats to Business Risk

• 

  Orchestrating Parallel Investigations: Inside the AI Swarm

The Real-World Friction

In modern enterprise environments, Security Operations Center (SOC) analysts suffer from acute alert fatigue. When Google Cloud Security Command Center (SCC) flags an anomaly—such as a compromised Service Account creating keys, escalating IAM permissions, or accessing Cloud Storage—an analyst spends 45–90 minutes manually querying Cloud Audit Logs, cross-referencing Cloud Asset Inventory, and tracing network flow logs. In an active data exfiltration breach, this manual dwell time results directly in data loss.

What it does

AegisFleet is an autonomous Tier 1 SOC agent fleet built natively on Google Cloud Platform.

• Asynchronous Ingestion: Subscribes to GCP Security Command Center findings via Cloud Pub/Sub.
• Parallel Swarm Investigation: Autonomous worker sub-agents (GCPAuditWorker, GCPAssetWorker, GCPIAMWorker) gather contextual audit logs and asset states concurrently.
• Gemini 3.5 Correlation Engine: Correlates disparate event streams to reconstruct the full attack chain, calculate blast radius, and detect privilege escalations.
• Automated Artifact Generation: Produces an unwatermarked visual Mermaid attack graph, a CISO briefing, and a prioritized containment checklist.
• Safe Human-in-the-Loop (HITL) Containment: Stages executable gcloud isolation commands in a sandbox. Upon one-click human authorization, the agent immediately revokes compromised keys and isolates targeted Cloud Storage buckets.

How we built it

We architected AegisFleet using the Google Antigravity SDK for the core agentic reasoning loop and deployed it as a serverless swarm on Google Cloud Run. The LeadOrchestrator agent is powered by Gemini 3.5 Pro to synthesize the raw telemetry. To maintain cross-session state across stateless Cloud Run instances, we implemented a persistent Memory Bank using Google Cloud Firestore. For data sanitization, we integrated Gemma 2B as a local pre-filter to redact sensitive customer PII from logs before they hit the main LLM context window. The backend is served via FastAPI with an interactive dashboard built using Tailwind CSS and Mermaid.js.

Challenges we ran into

• Infinite Tool-Call Loops: Early versions of the sub-agents would repeatedly poll Cloud Audit Logs with slight micro-variations if they couldn't find an event, exhausting API quotas. We solved this by implementing strict Antigravity Decide Hooks to act as circuit breakers and enforcing execution turn limits.
• Stateless Amnesia: Cloud Run scaling down to zero meant the agent forgot the incident context while waiting for human authorization. We overcame this by serializing the agent's execution graph and checkpointing it to Firestore.
• Indirect Prompt Injection: Feeding raw, untrusted security logs directly into the LLM created vulnerabilities where an attacker could embed commands in user-agent strings. We implemented strict XML quarantine delimiters (<untrusted_telemetry>) and Google Model Armor to block these injections.

Accomplishments that we're proud of

We successfully compressed a 60-minute manual triage process into 10 seconds of automated, highly accurate action. More importantly, we achieved strict deterministic governance: decoupling the reasoning engine from the execution environment ensures the autonomous swarm cannot perform destructive cloud mutations without authenticated human approval.

What we learned

Building robust AI agents requires shifting from prompt engineering to defensive systems engineering. Relying on strict Pydantic V2 schemas for tool inputs and outputs is the only way to prevent LLM hallucination and schema drift in production environments.

What's next for AegisFleet

We plan to expand the agent registry to support multi-cloud correlations across AWS GuardDuty and Azure Sentinel. Additionally, we are building direct Slack and Microsoft Teams webhook integrations so SOC analysts can authorize zero-trust containment directly from their mobile devices.

Built With

• antigravity-sdk
• fastapi
• gemini-3.5-pro
• gemma-2b
• google-cloud-firestore
• google-cloud-pubsub
• google-cloud-run
• google-security-command-center
• mermaid.js
• pydantic
• python