Skip to content

Security Model

Argus is designed to be safe to deploy in production environments. This page documents the security boundaries and data handling.

Read-only access

Argus uses read-only IAM roles on every cloud. It cannot modify, delete, or create any resources in your account. The specific permissions are documented on the IAM Permissions page.

Cloud Auth method Scope
AWS IAM execution role (Lambda) or assumed role (multi-account) Read-only: Resource Explorer, CloudWatch, Cost Explorer, CloudTrail, Describe APIs
GCP Service account with viewer roles Read-only: Asset Inventory, Monitoring, Logging, BigQuery
Azure System-assigned managed identity Read-only: Resource Graph, Monitor, Cost Management, Activity Log

Data collected

During a scan, Argus reads:

  • Resource metadata — IDs, types, regions, tags, creation dates
  • Usage metrics — CPU, network, disk I/O, connections (aggregated, not per-request)
  • Cost data — estimated monthly cost per resource in USD
  • Activity timestamps — when each resource was last accessed or modified

Argus does not read:

  • File contents, database records, or application data
  • Network traffic or request logs
  • Secrets, keys, or credentials stored in your resources
  • PII or customer data

Data flow

Cloud APIs → Argus agent (in-memory) → AI provider → Report
                                                        ↓
                                              Storage (S3/GCS/Blob)
                                                        ↓
                                                  Slack digest
  1. Resource data is fetched from cloud APIs and held in memory during the scan
  2. A compressed summary is sent to the AI provider for analysis (one batched call)
  3. The AI returns findings with waste reasons and recommendations
  4. A JSON + HTML report is saved to cloud storage (if configured)
  5. A compact Slack digest is posted with a link to the full report
  6. All in-memory data is discarded when the function exits

Data retention

Location Retention Encryption
Cloud storage (S3/GCS/Blob) 90 days (lifecycle rule on S3; configurable on GCS/Blob) Encrypted at rest (cloud-native SSE)
Pre-signed/SAS URLs 7 days (configurable via REPORT_URL_EXPIRY) HTTPS in transit
Slack messages Controlled by your Slack workspace retention policy Slack's encryption
Lambda/Cloud Run/Function memory Ephemeral — discarded after each invocation N/A

AI provider data handling

The AI call sends a compressed resource summary (IDs, metrics, costs) — not raw cloud API responses. No credentials, secrets, or application data are included.

Provider Data path Retention
AWS Bedrock Stays in your AWS account. Model invocation logs are off by default. No training on your data (AWS policy)
Anthropic API Sent to Anthropic's API endpoint. Not used for training. See Anthropic's privacy policy
Vertex AI / Azure OpenAI Stays in your cloud account Governed by your cloud provider agreement

Credential handling

  • No hardcoded credentials — all auth is via environment variables or cloud-native IAM roles
  • Multi-account STS sessions — temporary credentials, 1-hour expiry, never stored to disk
  • Webhook URLs — marked NoEcho (AWS) / @secure() (Azure) in deploy templates
  • API keys — stored in environment variables, never logged

Network access

Argus makes outbound HTTPS calls to:

  1. Your cloud provider's APIs (same account/project/subscription)
  2. The configured AI provider endpoint
  3. Slack's incoming webhook endpoint
  4. Cloud storage for report upload

No inbound network access is required. The Lambda/Cloud Run/Function does not expose any HTTP endpoints.