OpenAI expands agent tooling

OpenAI has added built-in sandboxing and a model-native harness to its Agents Software Development Kit, widening the tools companies can use to build autonomous software workers. (openai.com) The update landed on April 15, 2026. OpenAI said the new harness gives agents a standard way to work across files, tools, approvals, tracing, handoffs, and resumable tasks, while the sandbox provides an isolated place to run commands and generate artifacts. (openai.com) In plain terms, the harness is the control layer and the sandbox is the locked workroom. OpenAI’s developer docs say sandbox agents can manipulate files, install packages, expose ports, mount a data room, and continue stateful work later. (developers.openai.com) The first release is limited to the Python Agents Software Development Kit. OpenAI’s docs and outside coverage both say TypeScript support is planned for a later release. (developers.openai.com) (techcrunch.com) OpenAI framed the change as a way to move agent building away from custom glue code and toward a standard runtime for longer jobs. Its Agents Software Development Kit docs describe agents as applications that plan, call tools, hand work to specialists, and keep enough state to finish multi-step tasks. (openai.com) (developers.openai.com) The release comes as OpenAI is also widening access to stronger cyber models. On April 14, 2026, the company said it was expanding its Trusted Access for Cyber program to thousands of verified individual defenders and hundreds of teams, starting with a GPT-5.4-Cyber variant tuned for defensive security work. (openai.com) OpenAI said GPT-5.4-Cyber is available only to vetted security professionals, not the general public. The company said it is preparing for “more capable models” in the next few months and is adding safeguards as cybersecurity performance improves. (openai.com) The security push has also included a cleanup after a separate supply-chain problem. On April 10, 2026, OpenAI said a third-party developer tool called Axios was involved in a broader industry incident affecting the process used to certify its macOS applications as legitimate OpenAI software. (openai.com) OpenAI said it found no evidence that user data was accessed, that its systems or intellectual property were compromised, or that any software was altered. The company said older macOS versions of ChatGPT, Codex, Sora, and OpenAI FM would stop working after May 8, 2026, and asked users to update. (openai.com) Taken together, the April 2026 announcements show OpenAI adding more power for enterprise agents while tightening the controls around where that power runs and who gets access to its strongest cyber tools. (openai.com 1) (openai.com 2)