OpenAI Enhances AI Agent Security with Improved SDK

OpenAI Strengthens AI Agent Security with Enhanced SDK

An Enriched SDK for More Secure AI Agents

OpenAI has announced a significant update to its Agents SDK, an open-source development kit designed for creating autonomous AI agents. This improved version introduces two major new features to meet the growing security needs of businesses. The goal is to enable agents to execute long and complex tasks while ensuring the security of the systems on which they operate.

The first addition is a native harness for orchestration, specifically designed to integrate with OpenAI's advanced models. This component provides a comprehensive framework for agents, including configurable instructions, tools via the MCP protocol, persistent memory, shell command execution, and file editing. This orchestration aims to enhance the reliability of agents on complex tasks.

Integration of Sandbox for Enhanced Security

The second new feature is the native integration of sandbox execution. This functionality allows developers to run their agents in isolated environments, limiting their access to strictly necessary files, dependencies, and tools. OpenAI offers integration with seven sandbox providers, including Blaxel, Cloudflare, Daytona, E2B, Modal, Runloop, and Vercel.

Understanding the SDK Tools

• SDK (Software Development Kit): a set of software tools for developing applications, in this case, AI agents.
• Harness: an orchestration layer that manages the agent's instructions, tools, memory, and task coordination.
• Sandbox: an isolated execution environment for safely manipulating files and executing code.

To facilitate portability between different providers, the SDK introduces an abstraction system called Manifest. This allows for the description of the agent's workspace and its deployment across various compatible cloud services, such as AWS S3, Google Cloud Storage, Azure Blob Storage, and Cloudflare R2.

Towards More Robust and Scalable Agents

The separation between the "harness" and the "compute" is a central element of this update. By isolating the execution environment, developers can protect their credentials and sensitive data. This architecture also offers better durability, allowing for the saving and restoration of the agent's state in case of failure.

In terms of scalability, agents can now use multiple sandboxes simultaneously, parallelize tasks across multiple containers, or launch sub-agents in dedicated environments. This reduces the infrastructure work needed to transition from a prototype to a production deployment.

Transition to New APIs

OpenAI has announced the deprecation of its Assistants API, which will gradually be replaced by the Responses API and the Agents SDK by mid-2026. Current users of the Assistants API are encouraged to prepare for their migration.

A Strengthened Security Strategy

This update is part of OpenAI's strategy to structure its agent ecosystem. Since the experimental project Swarm launched in late 2024, which became a production tool in March 2025 with the addition of key components such as agent management, transfers between agents, safeguards, and traceability, OpenAI has continued to add key features for agent management and security. OpenAI has also launched the Responses API, introduced AgentKit during Dev Day, and launched Frontier to orchestrate agents in enterprises. OpenAI is working on additional capabilities, including a code mode and sub-agents, which will be deployed in both Python and TypeScript. However, CERT-FR advises against deploying autonomous AI agents in production within businesses. The Agents SDK is now available in Python, with TypeScript support coming soon, and is accessible to all API clients at standard rates.