Grab Builds Secure Agentic AI Workload Platform

In a significant advancement for the field of artificial intelligence, Grab has unveiled Palana, a proprietary platform designed to securely manage autonomous AI workloads. Developed by Grab's cybersecurity and platform engineering teams, Palana is a Kubernetes-native execution platform that provides a secure and isolated runtime environment. This innovative platform implements deterministic guardrails to mitigate the unpredictable behaviors commonly associated with model-driven applications, which are integral to the functioning of AI agents.

The creation of Palana stemmed from Grab's experiences while testing various agent frameworks, including OpenClaw. Through these prototype environments, the team recognized the necessity for a systematic, infrastructure-level approach to address the security risks posed by highly autonomous agents. Unlike traditional applications that operate within predictable parameters, model-driven agents possess the ability to execute arbitrary tools, interact with application programming interfaces (APIs), and manipulate source code independently. This operational autonomy, while powerful, introduces significant security vulnerabilities, including risks such as prompt injection, logic hijacking, dependency compromises, and hallucinations.

Security Framework and Isolation

At the core of Palana's design is the principle of isolation, which serves as the primary unit of trust within the platform. Adopting a zero-trust security model, Palana ensures that a security breach in one agent framework does not compromise adjacent workloads or the underlying compute cluster. Each agent is assigned to its own dedicated Kubernetes namespace, which is configured with stringent Role-Based Access Control (RBAC), custom network policies, and isolated service accounts. This meticulous configuration allows for enhanced security and operational integrity, as agents are provided with persistent, localized storage that preserves their state and memory across container restarts during extended asynchronous workflows.

Traditional methods of credential management, such as passing sensitive information through environment variables or mounted files, are deemed too risky for autonomous agents. A compromised runtime could potentially expose high-value API keys, leading to severe security breaches. To counter this risk, Palana employs a decoupled approach to secrets management, distinguishing between agent-readable credentials and proxy-only secrets. Sensitive credentials, including personal access tokens for version control and model gateway API keys, are securely stored within HashiCorp Vault, ensuring that they remain protected from unauthorized access.

When an agent needs to make an outbound API call, it is provisioned with abstract, dummy placeholder tokens instead of actual secrets. A secure intermediate proxy intercepts the request, validates the destination, and dynamically substitutes the placeholder with the real secret. This method ensures that the raw secret is never exposed within the agent container's environment, execution memory, or log files, thereby significantly reducing the risk of data leakage.

Centralized Security Controls and Operational Integrity

Given that agents must interact with external tools and model endpoints to function effectively, Palana's egress pathway is designed as a centralized security control point. All outbound HTTP and HTTPS traffic is routed through an Envoy proxy and an external authorization service that utilizes Open Policy Agent rules. This setup employs Man-in-the-Middle certificate authority termination, allowing for real-time decryption of traffic. Consequently, the platform can evaluate headers, validate endpoints, and perform token substitutions while generating comprehensive audit trails for security monitoring.

Recognizing that a compromised agent cannot be relied upon to terminate itself, Palana incorporates operational controls that exist outside the execution runtime. Network-level kill switches can disable network policies directly from the control plane, and an independent external reaper is responsible for triggering idle shutdowns without altering the core agent code. This layered approach to security ensures that even in the event of a breach, the integrity of the overall system remains intact.

The integration of Palana into a Kubernetes-native framework allows platform architects to leverage standard infrastructure-as-code practices for scaling and debugging agent runtimes. Each agent is treated as a custom resource, managed by a custom Kubernetes operator that dynamically provisions namespaces, storage, network policies, and ingress paths. This design bifurcates the operational experience, providing a simplified user interface and command-line tool for developers, while offering a robust, standardized Kubernetes layer for systems engineers. Platform teams can utilize these native definitions to programmatically audit, update, and manage the lifecycle of numerous concurrent agent workloads across the production cluster.

Why it matters

The introduction of Palana by Grab represents a pivotal moment in the ongoing evolution of secure AI workloads. By addressing the unique challenges posed by autonomous agents, this platform not only enhances security but also streamlines the operational management of AI systems. For creators and technologists, the implications are profound. The ability to run AI agents in a secure, isolated environment opens new avenues for innovation while minimizing the risks associated with deploying autonomous technologies. As organizations increasingly rely on AI-driven solutions, platforms like Palana could set new standards for security and operational integrity in the industry, ensuring that the benefits of AI can be harnessed without compromising safety or trust.