AI & Automation Briefing - October 3, 2026
Apple Tightens macOS Full Disk Access Controls Amid AI Agent Security Concerns
Apple announced it will add stricter controls around the macOS Full Disk Access setting, citing growing risks posed by AI agents. The setting was originally designed to support backup software, but it also grants apps permission to read files, mail, messages, and browsing history. Apple acknowledged in a developer blog post that some apps are using this access in ways users may not fully understand. The announcement followed a report from a journalist who claimed Meta's Muse app on Mac read his private messages without explicit permission, a claim Meta disputed, and a separate Wired report about a vulnerability in ChatGPT's Mac app. Going forward, Apple says it will require very explicit user action before any app can be granted this level of access. Apple stated that as AI agents become more capable and autonomous, the risks tied to Full Disk Access will continue to grow.
AWS Open Sources a Decision Model Built for Agentic Workflows
Amazon Web Services released Strands Decider 2B, an open source decision model designed to replace full LLMs at individual workflow steps where only routing or selection is needed. The model is built on a Qwen3.5-2B base but outputs calibrated confidence scores across a closed set of predefined options rather than generating text. It runs locally, operates at low latency, and costs less than frontier models to run. AWS distinguished engineer Marc Brooker developed the project after studying TypeSafe's Jev model, which pioneered this category. Brooker said AWS customers flagged the need after finding that full LLMs were overkill for many decision points inside agentic workflows. The release comes the same week OpenAI announced a comparable offering, reflecting rapid growth in the decision model category since TypeSafe introduced Jev. Strands Decider 2B reached the top spot on the Jevbench leaderboard for models its size before AWS formalized and released it through Strands Labs, the company's agent tooling division. The core engineering challenge, according to Brooker, is pushing accuracy and calibration without degrading the general language understanding that makes the model adaptable across use cases.
JevSpawn Enables Training-Free Agentic Inference via Compositional Action Spaces
Researchers have released JevSpawn, a training-free framework for agentic inference that builds action spaces directly from natural-language task descriptions. The agent explores possible actions using finite probability distributions and adjusts its behavior based on feedback from execution results. Shared context and structure across decision branches reduce redundant computation. Code and an interactive demo are publicly available.