OpenAI published a technical post detailing the Codex sandbox, a defense-in-depth layer that constrains the commands its coding agent runs locally through user-controlled filesystem and network limits. The team says it aims to make sandboxing easier to configure and to leverage stronger OS-level protections, expanding agent capability while keeping humans in control of access.
Defense-in-Depth
深度防御
Security principle layering multiple independent controls to reduce single-point failure risk.
Curated Intelligence
The Claude Code team disclosed their internal 'AI reviewing AI' workflow, which significantly boosted development efficiency through multi-layered checks. With approximately 90% of code generated by AI, merged changes doubled between June and September, while rollback rates dropped by over 40%. The system utilizes a 'Swiss Cheese Model' approach, shifting human focus from line-by-line reading to high-level judgment and accountability.
Proposes a four-layer defense-in-depth architecture for securing Model Context Protocol (MCP) in production, moving beyond simple gateway authentication. The framework addresses critical vulnerabilities like SSRF and schema drift by enforcing controls across secure tool execution, isolated management planes, restricted outbound trust, and semantic integrity via manifest pinning.
OpenAI research reveals that frontier reasoning models struggle to intentionally hide or distort their true thought processes within Chain-of-Thought (CoT) traces. This inherent limitation enables effective safety monitoring via CoT analysis and will be included in evaluation standards for future models like GPT-5.4 Thinking.