Security, in plain language.

Everything a bank security team needs to evaluate Astra: our controls, our sub-processors, where to send a vulnerability report, and how we respond when something goes wrong.

Report a vulnerability

security@codeastra.dev

PGP key available on request. Acknowledgement within 24 hours. Safe-harbor for good-faith research.

Email security

Controls

What we do, end to end.

Encryption

TLS 1.3 in transit. AES-256 at rest. Customer-held KMS keys (AWS KMS / Azure Key Vault) on roadmap for enterprise tier.

Deployment isolation

Cloud (multi-tenant), Hybrid (local vault + cloud LLM), and full Air-Gap Docker. Vault never co-locates with reasoning model.

Access control

Per-user, per-token RLS at the database layer. ES256-signed agent credentials. Master Admin keys are tenant-scoped.

Audit logging

Hash-chained, append-only. Every prompt, token resolution, and executor action recorded. Exportable in JSON or CSV.

Incident response

Acknowledgement within 24 hours of report. Initial assessment within 72 hours. Customer notification per contract.

Sub-processors

Every vendor that may touch your data.

On-prem and air-gap deployments call none of the sub-processors below. Cloud-tier customers can disable any LLM provider at the policy layer.

  • Supabase (AWS us-east-1)
    • Managed Postgres, auth, edge functions
      馃嚭馃嚫 US-East
  • Cloudflare
    • DNS, edge CDN, DDoS protection
      馃寪 Global
  • Anthropic
    • LLM inference (opt-in, no-training header sent)
      馃嚭馃嚫 US
  • OpenAI
    • LLM inference (opt-in, zero-retention API)
      馃嚭馃嚫 US
  • Google Vertex AI
    • LLM inference (opt-in, EU/US locality available)
      馃嚜馃嚭 EU / 馃嚭馃嚫 US
  • Stripe
    • Billing and subscription management
      馃嚭馃嚫 US / 馃嚠馃嚜 EU
  • Resend
    • Transactional email (notify.codeastra.dev)
      馃嚭馃嚫 US

Material changes to this list are communicated to enterprise customers at least 30 days in advance.

Data handling

What we promise. In writing.

  • We never train any LLM on your data. We send no-training headers (e.g. anthropic-beta: no-training-1) on every upstream call.
  • Sensitive values are tokenized before leaving your VPC. The reasoning model only sees opaque tokens like cdt_a3f9.
  • Zero-logging fallback: when enabled, prompts and responses are not persisted by Astra at all.
  • Customer data is never used for product analytics, fine-tuning, or evaluation without explicit written consent.
  • Data deletion requests are honored within 30 days. Audit logs may be retained per your contractual recordkeeping window.
  • On-prem and air-gap modes ship as a self-built Docker image; no telemetry leaves your perimeter.

Documents

Available on request.

Need a signed DPA or full questionnaire? We turn them around in 48 hours.