Grab Launches Palana, a Kubernetes-Native Platform for Securing Autonomous AI Agents

Grab's engineering and security teams have released Palana, a Kubernetes-native execution platform designed to safely isolate and govern autonomous AI agents in production. The platform addresses a growing operational reality: while traditional software is predictable, agents powered by large language models can act arbitrarily, call external APIs, modify code, and pursue goals in unpredictable ways—creating substantial security risks around prompt injection, logic hijacking, and unauthorized access.

Zero-Trust Isolation at Scale#

Palana's core innovation is treating isolation as the primary security boundary. Each agent runs in its own Kubernetes namespace with restrictive role-based access controls, network policies, and dedicated service accounts. This approach ensures that a compromised agent cannot spread to neighboring workloads or the host cluster. Agents also receive isolated, persistent storage for maintaining state across long-running asynchronous tasks.

Secrets Without Exposure#

The platform decouples credential handling entirely from agent execution. Rather than passing API keys as environment variables—a practice that would expose raw secrets if the agent's runtime were compromised—Palana stores sensitive credentials in HashiCorp Vault. Agents receive only abstract placeholder tokens. When an agent makes an outbound API call, a centralized proxy intercepts the request, validates the destination, and substitutes the real secret dynamically. The raw credential never touches the agent container, its memory, or its logs.

Centralized Traffic Control#

All outbound traffic flows through an Envoy proxy and Open Policy Agent rules, enabling real-time inspection and token replacement. The platform also implements network-level kill switches and external reaper processes—operational controls that exist outside the agent runtime itself, ensuring that even a fully compromised agent cannot evade shutdown.

Infrastructure-as-Code Operations#

Palana integrates natively with Kubernetes, allowing platform teams to manage hundreds of concurrent agent workloads using standard infrastructure definitions and custom operators. This separates the developer experience (simplified UI and CLI) from the systems engineering layer (standard Kubernetes resources), enabling programmatic lifecycle management and auditing across production clusters.