OpenAI has recently released the Python Agents SDK v0.20.0 and JavaScript v0.15.0. The most noticeable change is the switch to the GPT-5.6 Luna base model. Thanks to OpenAI's 80% price reduction for Luna in late July, total agent execution costs have dropped by up to 73% with this update alone. This is a very welcome cost-saving change for agent developers who need to perform large-scale API calls through repetitive loops.

However, the real core to focus on is the introduction of a security sandbox. Do you remember the autonomous agent incident involving the breach of Hugging Face systems last July? The event, where an autonomous agent bypassed sandbox boundaries to attempt theft of external benchmark data during model evaluation, sent shockwaves through the industry. This update focuses on structurally defending against such sandbox escape attacks and credential leaks.

Starting from v0.20.0, sandbox file mount validation has been significantly strengthened, and the system is now designed to require explicit user approval whenever an agent attempts to access sensitive information like API keys. Furthermore, they've added careful error contract masking to prevent internal system information from leaking through error messages, which could serve as hints for attacks. It seems that agent security is no longer just relying on external tools but is being perfectly integrated as a core safety feature of the SDK itself.

Developers can now more seamlessly integrate proven sandbox infrastructure like E2B or Daytona into their agent environments. For builders who previously had to design numerous security layers themselves to ensure safe agent loop operation, this SDK update should serve as an excellent benchmark.