Architecture
Isolated per-customer data
Your agent's conversations, memory, and files live in their own isolated data store — never mingled with other customers' data in shared tables.
Your choice of runtime
Your agent's hands-on tools (shell, files, browser) can run in our cloud, on your own computer, or on a dedicated machine you control.
Encrypted vault
Credentials, API keys, and tokens are stored in a vault encrypted with XChaCha20-Poly1305 and a key derived by Argon2id from your passphrase. Secrets are injected only at execution time — the model never sees them.
Vault execution boundary
Vault placeholders resolve only for eligible tool execution, and resolved values are removed from tool results before work continues. Ordinary prompt and tool content is not covered by a general sensitive-data scrubber.
Bounded tool access
Per-agent permissions gate shell, file, and browser access, and a command validator blocks dangerous patterns (destructive flags, pipe-to-shell installs, etc.) before any shell command runs.
Layered runaway controls
Search deduplication, loop detection, selected call budgets, supported write-action deduplication, and action caps reduce repeated or excessive tool use. Controls vary by tool and configuration.
Full audit trail
Every run, every step, every tool call is logged in a queryable journal so you can reconstruct exactly what your agent did and why.
Encryption in transit
Brainmox service endpoints use HTTPS/TLS for data in transit. This does not make third-party accounts or customer devices part of the same security boundary.
How we run
our own systems.
Found a vulnerability?
Tell us first.
If you believe you've found a security issue in Brainmox or the CloudAid infrastructure, email hello@brainmox.com with a description, reproduction steps, and your preferred contact. Please avoid including real customer data and allow a reasonable window for triage before public disclosure.