Security & Isolation
Every bot runs in its own isolated, sandboxed environment on shared hardware. This page describes exactly what that does and doesn’t protect against, so you can decide whether it’s the right fit for what you’re deploying — not as a marketing claim.
What isolates your bot
Section titled “What isolates your bot”- Its own private network, with nothing else attached to it. Two bots genuinely cannot reach each other over the network, even by IP — verified directly, not just asserted.
- Runs as an unprivileged, restricted process with no elevated system permissions and no way to acquire any — your code has nothing it would need root or a special capability for, and can’t get one even if it tried.
- A read-only filesystem, with a small, size-capped scratch space for incidental writes (a media attachment downloaded from the WhatsApp API before you process it, say). Nothing your code writes there survives a redeploy or restart, and nothing it writes can fill up shared storage.
- CPU, memory, and process limits, applied identically to every bot — one runaway or malicious process can’t starve the platform or other tenants of resources.
- No access to the underlying host — not its filesystem, not its other processes, not the systems that manage the platform itself. Your code cannot see or affect anything outside its own sandbox, at build time or at run time.
- A build timeout and an upload size limit — a build that hangs or an oversized upload fails cleanly instead of consuming resources indefinitely.
What this deliberately does not claim
Section titled “What this deliberately does not claim”- This is process-level isolation, not VM isolation. Every tenant shares the same underlying kernel on the same host. This is a real, meaningful boundary — not the same guarantee a hypervisor-level VM boundary provides. If your threat model specifically requires VM-level isolation, the current architecture doesn’t provide it.
- Outbound traffic shares an IP. Every bot on a given host currently sends outbound requests (including to Meta’s Graph API) from that host’s one public IP — there’s no dedicated per-bot egress IP yet.
- Not yet a multi-region or highly-available deployment. See Production limitations below.
Reporting abuse
Section titled “Reporting abuse”Every bot’s outbound traffic currently leaves through one shared IP (see above) — if you believe a bot hosted on Pywa Cloud is sending spam, scraping, attacking a service, or otherwise behaving abusively, email [email protected] with what you’re seeing (source IP/timestamps if you have them, and what the traffic looked like). Reports are the primary way a specific bad actor gets found and stopped without penalizing every other tenant sharing that IP.
Production limitations
Section titled “Production limitations”This is an early-stage platform. Being direct about what that means in practice:
- No independent, third-party security audit or penetration test has been performed. An internal adversarial review has (real findings, fixed - hardcoded fallback secrets that would have silently run in production if misconfigured, a signature-validation fail-open path, a missing role check on one endpoint, an unbounded-upload memory exhaustion path, among others) - a genuine pass, not a formality, but still the team building the platform reviewing its own work, not an outside assessment.
- Browser sessions are an
httpOnlycookie, with CSRF protection. The session cookie is scoped to the API host only, is invisible to JavaScript (so an XSS bug can’t read or exfiltrate it), and isSecure+SameSite=Lax. Any write made with it must also come from an allow-listed origin and carry a custom header, sinceSameSitealone doesn’t stop a sibling subdomain. The optional Pywa Studio sign-in uses the same cookie but is least-privileged: from Studio’s origin it can only read who is signed in and use Studio-specific endpoints, never change the account, manage teams, touch secrets or deploy tokens. Sessions slide: they expire after 7 days without use (renewed whenever you’re active) and can never outlive 30 days from the original sign-in. “Log out everywhere” invalidates every one of them. The one exception is the operator admin panel, which still keeps its single shared admin token inlocalStorage- a known gap, tied to the shared-credential item below. - The platform’s own backend has broader host access than any tenant’s sandbox does, by necessity of managing it. Further separating that from the tenant-isolation boundary described above is ongoing hardening work, not yet finished - the isolation guarantees above don’t depend on it, but ask if the distinction matters for your threat model.
- Admin operations currently use a single shared operator credential, not per-person accounts. Fine for a single operator; if you’re evaluating this as an organization with multiple support/ops staff, ask about per-admin audit trails before relying on them.
- No independent per-tenant network egress control or outbound-abuse-detection system exists yet beyond basic per-tenant rate limiting on inbound traffic.
- Backups exist and have been tested, but a specific production deployment’s own backup/restore procedure should be verified against that deployment’s real data before being relied on.
If any of the above is a blocker for your use case, ask before deploying anything sensitive — this page will be updated as the architecture changes, not left to go stale while the platform evolves.