Skip to main content
Phala Cloud runs customer workloads inside confidential virtual machines on hardware we operate. This page states the patching timelines we hold ourselves to, how a finding is classified, and how compliance is measured.

Scope

Customer workloads inside CVMs are out of scope. Phala does not scan tenant root filesystems. A customer’s own dependency posture is theirs to manage; we are responsible for the host, the image we publish and the platform services we run.

Severity

Severity is the CVSS base score of the finding, overridden upward in two cases: A finding with no available fix is still classified, but the remediation clock runs against a compensating control rather than a patch. See Exceptions.

Remediation timelines

The clock starts at first detection, not at first acknowledgement. Detection means the finding appeared in a scheduled scan, a vendor advisory we monitor, or a report we received.

What “remediated” means

A finding is closed when one of these is true and recorded:
  1. Patched — the fixed version is deployed to every affected asset.
  2. Mitigated — a compensating control reduces the exposure, with the control named and tested.
  3. Accepted — an owner has signed off with a written rationale and an expiry date no later than 12 months out.
For guest images, patching is not complete until the rebuilt image is re-measured and the new measurement is admitted by the allowlist. A patched image that cannot launch is not remediated.

Actively exploited vulnerabilities

On KEV listing or credible exploitation reporting, the on-call engineer opens an incident, not a ticket. Triage within 24 hours establishes whether Phala is affected and whether a mitigation exists that does not require a rebuild. Where the fix is a host kernel or firmware update, mitigation may be network-level isolation or disabling the affected feature; the patch follows on the normal track and the mitigation is recorded against the finding.

Exceptions

Some updates cannot be applied on the standard clock. The documented class is host kernel, NVIDIA driver and firmware updates on TDX hosts: applying them requires evacuating CVMs, and a driver update can break confidential-computing mode. An exception requires an owner, a compensating control, and an expiry date. It is reviewed at expiry and either renewed with justification or closed. An exception without an expiry date is not an exception.

Measurement

  • Scanning. Hosts and images are scanned with Trivy on a weekly cadence; container images on every build. Repositories use Dependabot.
  • Tracking. Every finding above Low enters a register with first_detected, severity, asset, owner, due_date and status.
  • Reporting. Overdue counts by severity are reported monthly. A finding is overdue when status is not closed and due_date has passed.
  • Enforcement. New production deployments pass a CI gate that fails on unresolved Critical findings. Until that gate is live, a named reviewer approves production endpoint changes.
The scan tooling and interpretation notes live in an internal repository under security/.

Current state

This policy is published while the platform is still catching up after the November 2026 incident. Two things are true and worth stating plainly:
  • The backlog is real. Hosts carry pending security updates accumulated before this policy existed. Those findings are being worked off and counted as overdue rather than grandfathered silently.
  • Guest image OS coverage is incomplete. Neither image line currently ships a package database, so scanner coverage of the guest OS is binary-level only until a build-time manifest lands (Dstack-TEE/dstack#1634).
We publish these gaps rather than describe the process as if it were already complete.

Reporting a vulnerability

Send reports to security@phala.network. For incident-specific matters, including key migration, use cert@phala.network. Please do not send secrets, private keys or customer data.