vArmor  by bytedance

Cloud-native container sandbox for AI workloads

Created 2 years ago
464 stars

Top 65.2% on SourcePulse

GitHubView on GitHub
Project Summary

Summary

vArmor is a cloud-native container hardening system designed to protect workloads, including AI Agents, within Kubernetes clusters. It leverages Linux kernel security features (AppArmor, BPF, Seccomp) and an Envoy-based network proxy to enforce granular access controls from system calls to application protocols. This enhances container isolation, reduces attack surfaces, and mitigates risks like container escape, lateral movement, and AI-induced prompt injection vulnerabilities.

How It Works

vArmor employs a multi-layered defense strategy by abstracting Linux's AppArmor LSM, BPF LSM, and Seccomp into distinct "enforcers." These can be combined with a sophisticated Envoy-based network proxy sidecar for L4/L7 egress control. This approach allows fine-grained policy enforcement on file access, process execution, and network traffic. The system's design prioritizes usability with an "Allow-by-Default" model, minimizing performance impact while offering robust auditing capabilities.

Quick Start & Requirements

vArmor follows a Kubernetes Operator pattern, enabling workload hardening via CRD API manipulation. Specific installation commands are not detailed, but official documentation, quick start guides, usage instructions, policies, and performance specifications are available via provided links. Prerequisites include a Kubernetes cluster; specific hardware or software version requirements beyond general Kubernetes compatibility are not explicitly stated.

Highlighted Details

  • Cloud-Native Design: Implemented as a Kubernetes Operator for seamless integration and management via CRDs.
  • Multiple Enforcers: Combines AppArmor, BPF, Seccomp, and an Envoy-based Network Proxy for comprehensive security controls.
  • Network Proxy Enforcer: Provides L4/L7/TLS SNI egress control, audit logging, and dynamic policy updates, extending Kubernetes NetworkPolicy.
  • AI Agent Protection: Offers defense-in-depth for AI workloads, mitigating prompt injection, tool abuse, and data exfiltration risks.
  • Flexible Policy Models: Primarily supports "Allow-by-Default" for usability and performance, but is also capable of "Deny-by-Default" using allowlist profiles.
  • Built-in Rules: Comes with pre-configured rules to simplify profile creation and deployment.

Maintenance & Community

Developed by ByteDance's Elkeid Team, vArmor is actively maintained and under development. Community engagement is facilitated through a Lark group. The project has also joined the 404Starlink initiative.

Licensing & Compatibility

The core vArmor project is licensed under Apache 2.0. However, its eBPF components are licensed under GPL-2.0, which may impose copyleft restrictions on derivative works. Users must comply with both licenses and those of any third-party components.

Limitations & Caveats

vArmor does not offer the same level of isolation as hardware virtualization solutions like Kata Containers. For scenarios demanding high-intensity isolation, users are advised to consider hardware virtualization containers alongside CNI's NetworkPolicy.

Health Check
Last Commit

9 hours ago

Responsiveness

Inactive

Pull Requests (30d)
15
Issues (30d)
0
Star History
9 stars in the last 30 days

Explore Similar Projects

Feedback? Help us improve.