Security
Trivy: vulnerabilities and misconfiguration
Install Trivy Operator and light up all four security lenses — CVEs and exposed secrets, misconfiguration, RBAC posture and the CIS benchmark.
On this page
Trivy Operator scans what is actually running: the CVEs in your images, workload misconfiguration, exposed secrets, RBAC posture and the CIS benchmark. Its findings land across all four lenses — Vulnerabilities, Configuration, Permissions and Compliance — of Security & Compliance.
Nothing is pushed. Trivy writes CRDs and KubeBolt lists them on its sweep — no token, no webhook, no exposed endpoint. KubeBolt does not scan anything itself: what it shows is what the operator reported.
What KubeBolt reads
Six report types, all aquasecurity.github.io/v1alpha1, listed cluster-wide on
every sweep:
| Report | Scope | Lens |
|---|---|---|
VulnerabilityReport | per workload | Vulnerabilities |
ExposedSecretReport | per workload | Vulnerabilities |
ConfigAuditReport | per workload | Configuration |
RbacAssessmentReport | per Role | Permissions |
ClusterRbacAssessmentReport | per ClusterRole | Permissions |
ClusterComplianceReport | cluster | Compliance |
The infra-assessment reports (InfraAssessmentReport,
ClusterInfraAssessmentReport) do not become findings. They are read only when
you open a compliance control, to name the resources that fail it.
Install
helm repo add aqua https://aquasecurity.github.io/helm-charts/
helm install trivy-operator aqua/trivy-operator \
-n trivy-system --create-namespace \
--set="trivy.ignoreUnfixed=true"
ignoreUnfixed=true is a recommendation, not a requirement. A CVE with no
published fix is real, but it is not a task — there is nothing to upgrade to.
Turn it off when you need the full inventory for an audit, and expect the list
to grow a lot.
The chart’s defaults already run every scanner KubeBolt reads. The compliance
reports — the CIS benchmark plus the other specs listed under
compliance.specs — run on their own schedule, compliance.cron, every six
hours by default. KubeBolt reads each report’s summary, which is what the
operator writes with its default compliance.reportType=summary; keep it.
How KubeBolt finds it
By the label app.kubernetes.io/name=trivy-operator, falling back to pods
named trivy-operator* in the trivy-system, trivy-operator or security
namespaces. A hand-rolled manifest without the standard labels is still found
if it lives in one of those. The result is the Trivy Operator card under
Administration → Agents & Ingest → Integrations, for the cluster selected in
the top bar.
Reading the reports needs a connection that can list them: direct, in-cluster, or the agent with its proxy on. A metrics-only agent produces no findings — see What it needs.
Verify
Reports appear a few minutes after install, one per workload:
kubectl get vulnerabilityreports -A
kubectl get exposedsecretreports -A
kubectl get configauditreports -A
kubectl get rbacassessmentreports -A
kubectl get clusterrbacassessmentreports
kubectl get clustercompliancereports
Findings show up in KubeBolt on the next sweep, within 10 minutes.
| Symptom | Cause |
|---|---|
| Card says “not installed” | No pods matching the label or the fallback namespaces |
| Card says “degraded” | Operator pods present but not Ready |
| Reports exist, nothing at all in KubeBolt | The cluster is connected metrics-only, or (self-hosted) authentication is off |
| Reports exist, no CVEs in KubeBolt | They are all MEDIUM/LOW — see below |
| No reports after 10 minutes | The operator cannot pull images (private registry, no credentials) |
| CIS empty, everything else fine | The compliance report runs on its own schedule (every six hours by default); give it a cycle |
| Permissions lens empty while RBAC reports exist | Every failing check sits on Kubernetes’ built-in roles, which are skipped |
| A fixed CVE never resolves | One of the six report types cannot be listed — its CRD is missing on an older operator, or the connection lacks permission. A finding is never resolved from a partial sweep |
What becomes a finding
CVEs: only CRITICAL and HIGH. MEDIUM, LOW and UNKNOWN are dropped
on purpose. On a busy cluster they are thousands of rows, and a list nobody can
finish is a list nobody reads. The full detail always stays in the reports
themselves.
Misconfiguration: CRITICAL, HIGH and MEDIUM. Wider than the CVE gate,
deliberately. Trivy’s MEDIUM band is where “runs as root”, “can escalate its
own privileges” and “seccomp disabled” live — several of those matter more than
things it calls HIGH. LOW is dropped: it is dominated by “runs with a low
UID” (10000 or below) and “CPU/memory not limited”, and the latter is a capacity
concern KubeBolt already raises as an insight. Each finding is
titled by its check id (AVD-KSV-…), so a reworded check does not reopen it.
Compliance: only failing controls. A passing control produces nothing. A
failing control keeps its severity when it is CRITICAL or HIGH, and is
filed as medium otherwise.
Exposed secrets are counted apart from CVEs, and they have no severity
gate: a LOW one is still a live credential in a layer anyone who can pull
the image can read. An unrecognised severity band is treated as HIGH. The
finding never contains the secret — Trivy’s partially masked match text is
not stored, only the file and the image — and the remediation says to rotate
the credential, not just remove it.
RBAC assessment: CRITICAL, HIGH and MEDIUM, the same gate as
misconfiguration, for both namespaced and cluster-wide roles. Kubernetes’ own
bootstrap roles — system:*, cluster-admin, admin, edit, view — are
skipped: they are identical on every conformant cluster, and the control plane
puts them back if you edit them, so a finding on one could never be acted on.
Each CVE carries its remediation ready to act on —
upgrade <pkg> <installed> → <fixed> — and the full image reference, rebuilt from the three fields
Trivy splits it across. That last part matters more than it looks: two
workloads sharing an image are one fix, and without the reference nothing
in the row says so.
Trivy reports against the ReplicaSet. KubeBolt files each finding under the owning Deployment, so a rollout does not reopen it with its age back at zero.
The CIS rollup trap
Some CIS controls are not independent checks — they are sums of checks
KubeBolt already stores one by one. Any control whose checks are AVD-KSV-*
is evaluated from the same config-audit reports that already produced their own
findings.
Counting both would file the same problem twice: once as N workload
misconfigurations, again as the control that groups them. Those controls are
flagged as rollups, so the dashboard shows them as posture without adding
them to the actionable total. Controls evaluated by control-plane and node
checks (AVD-KCV-*) are never rollups: they are the only place that class of
problem shows up.
This was verified against a live cluster: for all 35 controls where both sides had data, the control’s own failing count equalled the summed failing results of exactly those checks. The duplication is arithmetic, not approximate.
Cost of the sweep. KubeBolt lists six report types per cluster, each a cluster-wide LIST across the agent tunnel. The cost tracks payload, not object count: measured on a dev cluster, 57 vulnerability reports took 110 ms while 120 config-audit reports took 29 ms — a vulnerability report carries hundreds of CVEs, an assessment report a handful of checks. If the sweep ever shows up in your latency, the lever is Trivy’s scan scope, not KubeBolt’s read.