Security

What we do, what we don't, and what's not in scope yet.

Security practices for the platform and our cloud. Honest about scope — we'll add what we don't have as we have it.

Last updated: 2026-11-06. Security questions or vulnerability reports: use our contact form with the topic "Security issue".

Live Integrity — what we catch, what we don't

Live Integrity is our umbrella for two related rootkit-detection features: the scan (universal — out-of-band on our cloud, an in-guest cross-view on other providers) and the integrity watch (scans on a schedule, ComputeForge cloud only), both on a server's Integrity tab in the panel. Both exploit the same architectural fact, and both have honest limits.

The architectural insight, in plain English. Kernel rootkits hide things by hooking the read path inside the guest kernel — read() syscalls, /proc traversal, netfilter hooks, the ls and ps binaries themselves. Anything that reads the OS through those hooks reads what the rootkit wants it to read. That includes your in-VM antivirus, your auditd, your file-integrity-monitor agent — they're all downstream of the kernel that's lying to them. When ComputeForge reads from outside the guest — from a block-level copy of your disk, read by another machine — the read never traverses the guest kernel at all. The rootkit has no hook point. When the in-VM view and the outside view disagree, that difference is what we flag: it is how a hiding rootkit gives itself away.

Concretely:

ThreatScan on other providers (in-guest)Scan / integrity watch on our cloud (out-of-band disk)
User-space trojanspartial (changed package files, persistence entries)✓
LKM rootkits with on-disk modules✓✓
Patched kernel imagesfull scan onlyfull scan only (not the scheduled watch)
Files hidden by syscall hookspartial✓
Hidden processesnotnot (needs the memory view)
Hidden listening portsnotnot (needs the network view)
Memory-only rootkits with on-disk delivery vectorpartialpartial (the on-disk vector)
Memory-only rootkits with no disk trace, no network IOCnotnot
Virtualization-aware kernel rootkitspartialpartial (what persists on disk)

The honest limits. Memory-only rootkits with no on-disk delivery vector and no outbound network IOC are out of scope for both tools — there's nothing on the disk to find. Sophisticated nation-state implants can theoretically defeat any scanner: APT-grade tooling that knows it's being introspected can mimic baseline behavior. Live Integrity is calibrated for opportunistic and commodity threats — backdoored OpenSSH packages, cryptominers, persistence-focused commodity rootkits, supply-chain tampering. APT-grade adversaries are not our threat model, and we won't pretend they are.

Two further constraints worth naming. The package-hash checks cover current Ubuntu LTS, Debian stable, Rocky Linux and AlmaLinux, with best-effort coverage on other distros. And the integrity watch reads the disk only: network and memory views would need a read-only sensor on our host, which is not available yet.

Autonomous action policy Coming soon

Lifeguard can act on your VMs when an anomaly is detected — branch the box for forensics, fail traffic over to a standby via Forge DNS, check the compromised VM out to its last good tag. Whether we act, propose, or just alert is a policy you set per service. Three modes:

ModeCompromise detectedHardware failureDeployment regression
alert-onlyPage on-call. Capsule sealed. No state changes.Page on-call. Service stays down until human acts.Page on-call. No automatic rollback.
propose-and-approveCapsule sealed. Chain proposed in panel + paged. Human one-click approves.Failover to the standby proposed. Human approves.Rollback chain proposed (revert tag, re-route DNS). Human approves.
full-autonomousChain executes. Capsule sealed. You're notified after.Traffic fails over to the standby automatically. You're notified after.Auto-rollback to last known-good tag. You're notified after.

Per-service policy. The autonomy mode is set in your Forge Topology spec, per service. Production databases default to propose-and-approve (because we don't ever want to auto-act on stateful primary data without explicit human signoff). Stateless web tiers commonly run full-autonomous. Staging environments commonly run full-autonomous too — the whole point of staging is to absorb risk. Mixed-policy fleets are normal: prod-db = propose-and-approve, prod-web = full-autonomous, staging = full-autonomous.

Auditable. Every autonomous action is recorded in cf log with the trigger (anomaly type, signal values, baseline deviation), the decision (which chain, which policy gate evaluated), the executor (Lifeguard signing key + orchestrator commit ID), and the outcome (success / partial / rollback), and appears in your activity log, which exports as CSV or JSON-Lines through the API.

Revertable. Customer can revert any autonomous action with cf revert <action-id> — subject to physical constraints. We can restore a VM from its checkpoint while the checkpoint is kept; we cannot un-destroy a VM whose disks were already reclaimed. DNS swaps revert in seconds. A server taken out of service by a whole-server checkout is kept, stopped, for 24 hours, so it can be put back if the autonomous decision was wrong.

Mesh ACLs

Forge Mesh is a WireGuard-based overlay; a VM you add to it gets peer-to-peer connectivity to all other VMs on your account's mesh by default. ACLs are configurable in the panel as rules of which devices may reach which devices, on which ports. The mesh coordinator hands the rules to every peer.

Sharing with another organization is possible through the API and explicitly opt-in: the owning organization names the devices (or all of them) and the ports it shares. We do not bridge meshes implicitly.

Signing in

Two-step sign-in is available for the panel: an authenticator app, a passkey or an e-mailed code. An organization can require it for all its members: then an authenticator app or passkey is needed (an e-mailed code alone doesn't count), and API keys of members without it are refused until they set it up. SSH access to your servers uses the public keys you add in the panel or through the API.

Activity log

Changes made to your account through the panel or the API are recorded in your activity log: each change that succeeds (creating, changing or deleting a server, an SSH key, an API key or a member, and approvals), with who made it and when, plus requests that were refused. Reading data is not recorded. Some actions we take are not in it yet: of our automatic measures, only the shut-off of a server's outbound mail writes an entry. The log is kept for the life of the account and exports as CSV or JSON-Lines through the API.

Network

The panel and API are served over TLS only. Your servers' public addresses carry reverse DNS (PTR) records you can set: IPv4 addresses, and addresses in the IPv6 subnets you add to your account.

What's NOT in scope yet

Honest disclosure of where we are vs. where we're going:

  • SOC 2: we do not hold a SOC 2 report. The practices on this page are externally unaudited.
  • ISO 27001: not pursued at this time.
  • FedRAMP: not pursued. We are not currently a viable host for U.S. federal workloads.
  • HIPAA BAA: we do not sign Business Associate Agreements.
  • PCI DSS: we do not store cardholder data (our payment processor holds it). Customers running card-handling workloads on our VMs are responsible for their own PCI scope; we can provide architecture support.

Vulnerability disclosure

If you find a vulnerability, please tell us through our contact form with the topic "Security issue". We do not currently run a paid bug bounty.