Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB
CVE-2026-95837 Overview
CVE-2026-95837 affects containerd, an open-source container runtime widely used in Kubernetes and other orchestration platforms. The vulnerability resides in the Container Runtime Interface (CRI) CreateContainer checkpoint restore path. When restoring from a checkpoint, CRIU (Checkpoint/Restore In Userspace) rehydrates process security state from the checkpoint data rather than enforcing the destination ContainerConfig. An attacker who can supply a crafted checkpoint can restore a container that runs as root with full Linux capabilities, bypassed no_new_privs, and no seccomp filters, while the CRI API reports the originally requested, restricted configuration. The flaw is categorized under [CWE-269: Improper Privilege Management].
Critical Impact
Attackers with container creation privileges can escalate to full root with all capabilities inside containers while CRI status falsely reports a restricted security context.
Affected Products
- containerd versions 2.1.0 through 2.2.6
- containerd 2.3.0 through 2.3.3
- Systems using CRI checkpoint restore via CreateContainer
Discovery Timeline
- 2026-10-12 - CVE-2026-95837 published to NVD
- 2026-10-16 - Last updated in NVD database
Technical Details for CVE-2026-95837
Vulnerability Analysis
The containerd CRI implementation supports container checkpoint and restore operations using CRIU. During restore via CreateContainer, containerd should apply the security constraints defined in the destination ContainerConfig, including user ID, Linux capabilities, no_new_privs, and seccomp profile. Instead, CRIU reconstructs the process using security state embedded in the checkpoint image.
This means an attacker who crafts a checkpoint can embed arbitrary credentials and privilege settings. The restored process executes with those attacker-controlled values. The CRI status response continues to report the requested, restrictive configuration, hiding the true runtime state from orchestrators and monitoring tools. Systems that never invoke CRI checkpoint restore are not affected.
Root Cause
The root cause is a trust boundary failure between checkpoint data and the CRI container specification. The restore path treats checkpoint-embedded security state as authoritative instead of re-applying the policy from ContainerConfig. Capabilities, UID/GID, no_new_privs, and seccomp filters supplied by the restore source override those requested by the caller.
Attack Vector
Exploitation requires local access to a containerd host where CRI checkpoint restore is enabled, and the ability to submit a crafted checkpoint through CreateContainer. The attacker builds a checkpoint that encodes root UID 0, a full capability set, disabled no_new_privs, and an empty seccomp filter. On restore, the resulting container escapes the intended least-privilege sandbox and can be used as a staging point for container escape or lateral movement.
// Security patch in internal/cri/config/config.go
// cri: disable restore in CreateContainer by default
// EnableCRIU enables CRIU (Checkpoint/Restore In Userspace) support.
// When set to false, checkpoint/restore operations will be disabled.
EnableCRIU *bool `toml:"enable_criu" json:"enableCRIU"`
// EnableExperimentalRestoreViaCreate enables experimental restore of container
// checkpoints via CreateContainer.
// When set to false, checkpoint restore via CreateContainer will be disabled.
// Default: false
EnableExperimentalRestoreViaCreate bool `toml:"enable_experimental_restore_via_create" json:"enableExperimentalRestoreViaCreate"`
Source: containerd commit 928c03c6
Detection Methods for CVE-2026-95837
Indicators of Compromise
- Containers whose actual runtime UID, capability set, or seccomp mode diverges from the ContainerConfig reported by the CRI ContainerStatus API.
- CRI CreateContainer calls that reference a checkpoint image path or restore annotation on affected containerd versions.
- Unexpected processes running as UID 0 with CapEff containing capabilities such as CAP_SYS_ADMIN inside workloads provisioned as non-root.
Detection Strategies
- Compare pod spec security context against runtime process state by inspecting /proc/<pid>/status for Uid, CapEff, NoNewPrivs, and Seccomp on container PIDs.
- Audit containerd and Kubernetes API logs for CreateContainer requests carrying checkpoint or restore parameters.
- Flag any container restore performed on containerd 2.1.0 through 2.2.6 or 2.3.0 through 2.3.3 where CRIU is enabled.
Monitoring Recommendations
- Enable Kubernetes audit logging at RequestResponse level for pod create operations and alert on checkpoint-related annotations.
- Collect host-level process telemetry to detect capability or UID mismatches between declared and effective container security context.
- Monitor containerd configuration drift to ensure enable_experimental_restore_via_create remains disabled across the fleet.
How to Mitigate CVE-2026-95837
Immediate Actions Required
- Upgrade containerd to version 2.2.7 or 2.3.4, which disable experimental checkpoint restore via CreateContainer by default.
- Set enable_experimental_restore_via_create = false in containerd configuration on all nodes until upgrades are complete.
- Restrict access to the CRI socket (/run/containerd/containerd.sock) to trusted orchestrators and audit any client with CreateContainer privileges.
Patch Information
The fix is included in containerd 2.2.7 and 2.3.4. The patch introduces a new configuration flag EnableExperimentalRestoreViaCreate that defaults to false, disabling the vulnerable restore path. Review the GitHub Security Advisory GHSA-p7v4-vr35-mj6f and Pull Request #13913 for full remediation details.
Workarounds
- Disable CRIU support entirely by setting enable_criu = false where checkpoint/restore is not required.
- Block or reject CreateContainer requests containing checkpoint restore parameters at an admission controller or policy engine.
- Use pod security policies or OPA/Gatekeeper rules to deny workloads that attempt to run as root or request elevated capabilities when the workload spec declares otherwise.
# containerd configuration (/etc/containerd/config.toml)
# Disable experimental checkpoint restore via CreateContainer
[plugins."io.containerd.cri.v1.runtime"]
enable_experimental_restore_via_create = false
enable_criu = false
# Verify installed version
containerd --version
# Restart containerd to apply configuration
systemctl restart containerd
Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB

