Security

This page describes the security model as implemented. For the public commitments and how to report an issue, see /security and responsible disclosure.

Runtime isolation

Generated and user code never runs inside the web application. The runner service executes it in a sandbox:

• an unprivileged user inside Linux user, PID, mount and network namespaces,
• cgroup limits on memory and process count,
• wall-clock timeouts for every command,
• allow-listed commands only (install, build, test, typecheck, the template's dev command); the agent cannot run arbitrary shell,
• optional Docker and E2B drivers for operators who prefer container or remote isolation.

Previews are reached through a host-based proxy that maps a hostname to one sandbox; sandboxes cannot see each other.

Secrets

• Environment variable values, integration credentials and GitHub tokens are encrypted with the platform encryption key before they are stored.
• The API returns masked values; revealing a value requires the editor role and is logged.
• Secrets are never included in logs, AI context or tool outputs.
• Password hashes, session tokens and email tokens are stored hashed.

Sessions and authorization

• Sessions are httpOnly cookies backed by hashed tokens with expiry.
• Every API route authenticates the user and checks workspace role or project access on the server; client-supplied ids are never trusted.
• Admin pages require a staff account.
• Auth and AI endpoints are rate limited.

AI boundaries

• The agent only sees file contents, the file tree, recent messages and sanitized tool output; never environment variable values.
• Paths are validated so tools cannot escape the project.
• Runs are bounded in tool calls, turns, time and tokens, and start with a snapshot so changes are reversible.
• Prompts and file contents are sent to the configured provider (Anthropic or OpenAI). The AI policy describes this in user-facing terms.

Webhooks

Stripe and Razorpay webhooks are verified by signature and deduplicated by event id. See Webhooks.

What we do not claim

Codexa has no third-party certifications or audits (no SOC 2, ISO 27001, HIPAA or similar) at this stage. Statements about compliance programmes will only appear here when they are true.

Reporting a vulnerability

Email security@codexa.app or use the vulnerability reporting form.