Radar Daily Briefings

Mistral has released Shieldstral, a 3B open-weights multimodal safety classifier under Apache 2.0. It evaluates text, images, and text-image inputs using natural-language policy questions supplied at inference time.

The model formats each request as an instruction, a yes-or-no query, and a document to judge. It reads only the yes and no logits, then softmax-normalizes them into a continuous safety score. Mistral says the model supports moderation, refusal detection, and policy adaptation without retraining.

Mistral also says Shieldstral runs on a single 16GB NVIDIA GPU and matches or outperforms larger open guard models. The supplied release does not include detailed benchmark results or independent corroboration.

INTERPOL’s African Cyberthreat Assessment Report 2026 says artificial intelligence was involved in 55% of recorded cybercrime cases across Africa. The assessment covers data from 36 countries and identifies online scams as the most common cybercrime form in 2025.

According to INTERPOL, criminals are using AI with social media and mobile-money services, including for convincing business-email scams, deepfake-enabled sextortion, and synthetic identities. Reported cybercrime-related financial losses rose from $192 million in 2024 to $484 million.

The assessment also highlights scam centres, regional differences, and weak information sharing among banks, telecom companies, and law enforcement. The supplied evidence does not include the report’s methodology or independently validate its figures.

Liquid AI has released LFM2.5-2.6B, a 2.6-billion-parameter model intended for on-device agents. The model and its base version are available on Hugging Face, with guidance for local-agent harnesses and a WebGPU demo.

The training pipeline combines supervised fine-tuning, specialist-teacher distillation, and agentic reinforcement learning in real harnesses. Liquid AI reports support for llama.cpp, MLX, vLLM, SGLang, and ONNX, plus measured CPU and GPU throughput.

The release is relevant for engineers evaluating local tool use, multi-step workflows, and memory-constrained inference. Reported benchmarks and hardware speeds come from Liquid AI, and the article says larger models remain stronger for coding.

WordPress 7.1 introduces a unified meta.public flag for Abilities. It records whether an Ability is generally intended for external clients, including REST API consumers, MCP adapters, and AI agents.

For REST, show_in_rest takes precedence over public; otherwise public supplies the default, with absent values resolving to false. The resolved public property is always present in Ability metadata and is declared in the REST schema. Existing channel-specific registrations remain compatible.

The flag governs exposure and discoverability, not authorization. Permission callbacks remain mandatory, and developers should preserve explicit channel opt-outs. The WordPress MCP Adapter is slated to adopt the flag in its next release, while WP-CLI lists all registered Abilities.

PipeNetwork’s minimax-h3-mlx package ports MiniMax-H3 to MLX for running on Apple Silicon. Simon Willison tested it on an M5 Max MacBook Pro and generated a video from a text prompt.

MiniMax-H3 accepts text, images, audio, and video, and can generate video clips with audio. The experiment downloaded about 115 GB of model files, while video generation took just under 45 minutes.

The result demonstrates a demanding local multimodal workflow rather than lightweight inference. Willison reports that the audio was speech-like garbage because the prompt did not specify audio guidance; the supplied document does not establish broader performance or package lifecycle status.

Simon Willison has released LLM 0.32, a major update to the LLM CLI and Python library. It adds visible reasoning traces, server-side provider tools, OpenAI Responses API features, new models, and structured streaming events.

The Python API now accepts complete message histories through model.prompt(messages=[]), while stream_events() exposes reasoning, text, tool, and other event types. A content-addressable message store reduces duplicated conversation logging, and the llm openai endpoint command targets compatible APIs without logging those requests.

Tool chains can pause for human approval and resume from stored history. Plugins providing extra models need updates for full streaming-events support.

Nehir is a macOS tiling window manager built around Niri’s horizontally scrolling column layout. The supplied project documentation describes its capabilities and installation paths, but does not document a published release or released version.

Windows are arranged in columns that scroll horizontally, with overflow tabbing when stacked windows cannot fit at minimum heights. Nehir also provides workspaces, multi-monitor support, live-watched split TOML configuration, Unix-socket IPC through nehirctl, and application rules.

Developer Mode adds runtime dumps, state reset and restart actions, and trace capture through the command palette, shortcuts, or IPC. Users must grant Accessibility permissions, and force quits can leave parked windows at the screen edge.

Charlie Holland’s article explains Emacs Lisp macros as a metaprogramming mechanism built on the language’s homoiconic representation of code and data. It is an explanatory tutorial, with no release or project-status change described.

The article shows how macros receive unevaluated forms and generate code before execution, using quasiquotation, macroexpand-1, and examples such as unless and deftoggle. It also covers macrostep and related expansion commands for inspection.

For Elisp engineers, the guidance clarifies when macros are appropriate: controlling evaluation, establishing bindings, or defining new forms. The article cautions about repeated evaluation, variable capture, unreadable expansions, and recursive expansion; behavior may depend on the expansion environment.

The WordPress Test Team reports that its Test Handbook review process has concluded with the handbook merged into GitHub and officially synced. The post invites contributors to check it and report remaining problems.

The roundup also organizes active WordPress core and Gutenberg work into patch testing and bug reproduction. Contributors are asked to apply proposed patches, follow testing instructions, manually verify behavior, and submit reports.

The article is useful as a participation guide, with listed tickets, calls for testing, and August 2026 meeting times. The handbook may still contain bugs or presentation issues, and the supplied material does not include testing outcomes.