Apple gets a new CEO, DeepSeek Harness hits 112K stars, and other tech news
John Ternus takes over as Apple CEO ahead of the September event, DeepSeek releases its Harness agent framework, six CVEs land in ash_ai, plus OpenMAIC and the ArcBox isolation runtime.
Big Tech
John Ternus officially becomes Apple CEO
As Apple announced in April, John Ternus formally took over as CEO on 1 September. Tim Cook moves to Executive Chairman. Ternus joined Apple in 2001 and served as senior vice president of hardware engineering.
Apple has scheduled its annual product event for 9 September at Apple Park, under the tagline “Surprise and Shine.” It will be the first product launch led by a CEO who came up through hardware engineering.
Having overseen the iPhone, iPad, Mac, AirPods, and Apple Watch lines is his strength. The event is where the new leadership’s direction on both Apple Intelligence strategy and product hardware should become visible.
AI Security
Six vulnerabilities disclosed in the Elixir ash_ai LLM tool extension
Six security vulnerabilities were disclosed in ash_ai, an Elixir-based AI tool extension. Ash Framework maintainer zachdaniel published them together on 30 August. Versions from 0.1.0 up to (but not including) 1.0.0 are affected.
The most severe,
CVE-2026-77956,
is a remote code injection flaw. AshAi.Actions.Prompt evaluates prompts with EEx.eval_string/2,
letting an unauthenticated attacker execute arbitrary Elixir code on the server.
CVE-2026-81315 bypasses DNS rebinding protection in the MCP server. CVE-2026-82580 leaks internal information through error messages. The set is a reminder that security review of LLM tool chains is still at an early stage.
Open Source
DeepSeek Harness — an agent framework where everything is a plugin
DeepSeek Harness (dsh), the agent framework released by DeepSeek AI, is growing quickly on GitHub, passing 112K stars shortly after release. Its stated design philosophy is “Everything is a Plugin.”
Per the official site, models, tools, skills, sessions, sandboxes, storage, and UI are all plugins, replaceable and recombinable individually. Unlike existing frameworks, it modularizes down to the infrastructure layer.
It is currently a developer preview, with breaking changes expected. Handling context selection, state management, and permission control explicitly is what makes the approach notable in agent infrastructure.
OpenMAIC v1.0.0 — a multi-agent interactive classroom
OpenMAIC v1.0.0, developed by a Tsinghua University research team, shipped on 27 August. Give it a topic or a document and it generates a classroom experience complete with slides, quizzes, and simulations, where an AI teacher and AI classmates take part in live discussion.
v1.0.0 adds a Pro workbench for designing a curriculum in conversation with an agent and iterating on each page. It builds on the 3D visualization and in-browser programming introduced in v0.2.0 in April.
It supports a range of LLMs including OpenAI, Anthropic, Gemini, DeepSeek, and Ollama, and custom TTS/ASR providers can be plugged in. It is a concrete example of multi-agent systems applied to education.
ArcBox — a Rust-based isolation runtime for AI agents
ArcBox is an AI agent isolation runtime written in Rust, supporting macOS on Apple Silicon and Linux. Being OCI-compatible, it runs existing container images.
It offers three isolation tiers: Docker-compatible containers, Firecracker-based sandbox microVMs, and full VMs with a dedicated kernel. Sandbox cold boot is under 200ms and snapshot restore under 50ms.
It provides practical isolation for agent code execution and CI work, keeping kernel-level isolation while solving conventional VM boot latency. The target overhead is 10–30MB per agent.