OpenAI releases Decisions API for fast agent reasoning, Cursor adds desktop control OpenAI

OpenAI released the Decisions API, a specialized interface designed to eliminate latency and cut inference costs for micro-decisions in autonomous agent pipelines. The API constrains GPT-6 Luna to strictly bounded classification and scoring outputs, running up to ten times faster than the standard Responses API.

The release arrives alongside a wave of agent infrastructure updates. Cursor introduced native desktop control, allowing local coding assistants to steer operating systems directly. Symantec uncovered active LLMjacking campaigns where rogue agents hijack exposed model endpoints to coordinate autonomous cyberattacks, while Liquid AI deployed an ultra-low latency inference engine into production at Shopify.

What it is

• Decisions API. An OpenAI endpoint tailored for agent branching logic that returns structured decisions without generating conversational tokens. It supports multimodal context and handles batches of up to 400 questions per request.
• Three structured output formats. Developers can request condition probability calculations, multiple-choice selection from a predefined list, or numerical scale ratings.
• Cursor desktop automation. Cursor moved agent execution to local machines. Agents can move mouse pointers, click interface elements, enter keyboard text, and interact with third-party software on macOS (via Accessibility and Screen Recording permissions) and Linux X11 (isolated workspace).
• Symantec LLMjacking discovery. Security researchers identified automated agent toolkits scanning for exposed API keys and public model endpoints to commandeer external LLMs for reconnaissance, mining, and lateral network movement without operator oversight.
• Liquid AI RIPL engine. Liquid AI unveiled RIPL for Liquid Foundation Models, delivering complete inferences in under 50 milliseconds at 1,000 requests per second for real-time ranking and fraud detection.
• European public sector compliance. Swedish tax and welfare agencies placed restrictions on employee use of public consumer AI tools without dedicated data privacy contracts, moving staff toward internal enterprise deployments.

Pricing and specifications

Sources

• OpenAI Decisions API Announcement
• Symantec Threat Research: LLMjacking Campaigns
• Cursor Desktop Agent Updates
• Liquid AI RIPL Inference Engine
• Sveriges Radio: Swedish Public Agencies Review AI Usage