73 Microsoft Packages Poisoned in Attack Targeting AI Agents — A New Supply Chain Threat
Seventy-three open-source packages cryptographically signed by Microsoft have been found contaminated with credential-stealing malware. According to reports from multiple researchers at Step Security and OpenSourceMalware, the packages were rigged to launch a self-propagaging stealer the moment an AI coding agent opened them.
What's striking is how slow the response was. Although GitHub promptly disabled the packages, it treated the matter as a "terms of service violation" and never explicitly told users that the packages might contain malicious code. It wasn't until Monday—several days after the incident—that Microsoft headquarters took the repositories offline, saying it was "investigating potentially malicious content."
The incident is believed to be part of an attack campaign dubbed "Miasma Worm," and what's new here is that it exploits the way AI coding agents automatically load packages. Researchers warn that developers using AI agents should respond as though their systems have been compromised.
Apple Waives AI API Costs for Small Developers
As part of its WWDC-related announcements, Apple has rolled out a measure aimed at attracting small developers at a time when the cost of AI experimentation is soaring. Developers whose apps have fewer than 2 million first-time downloads on the App Store will get their cloud AI API costs waived.
The move is intended to lower the cost barrier facing developers who want to build AI features into their apps. It appears to be a strategy to let developers experiment freely as the model family underpinning Apple Intelligence is offered via the cloud.
At WWDC, the company announced "Siri AI" and unveiled "Apple Foundation Model," a multimodal model co-developed with Google; this free-tier policy can be read as a move to accelerate AI adoption across the entire ecosystem.
NVIDIA's LocateAnything-3B: Fast Visual Grounding via Parallel Box Decoding
NVIDIA has released LocateAnything-3B, a vision-language grounding model. Instead of conventional autoregressive coordinate generation, this 3B-parameter model adopts a technique called "Parallel Box Decoding" (PBD), achieving up to 2.5x higher throughput.
Rather than sequential token-by-token decoding, PBD predicts complete bounding box coordinates in a single parallel step. It is designed to improve inference efficiency while preserving geometric consistency, and runs on NVIDIA H100 GPUs.
The training data is a large multi-domain dataset comprising 12 million images, more than 138 million queries, and 785 million bounding boxes. It covers a wide range of domains, including natural scenes, robotics, autonomous driving, GUI interaction, and document understanding.
The model has been released for non-commercial use only under the NVIDIA License, and integration into Nemotron and Cosmos is also underway.
The jqwik Incident: An OSS Developer Planted a "Code-Deletion Trap" for AI
An incident surrounding jqwik, a property-based testing library for Java, has been drawing attention: an OSS developer deliberately planted a trap designed to make AI coding tools delete code.
According to an explanatory article on Qiita, the incident exposed a new attack surface in an era when AI generates and modifies code. It demonstrated that a developer can craft prompts or a codebase in such a way that AI agents are induced to unintentionally delete critical code.
As AI coding becomes commonplace, the incident is forcing a rethink not only of how to validate generated code, but also of the trustworthiness of the very context fed to AI.
Tool-Calling Benchmark for Qwen3.6-35B-A3B: ByteShape vs Unsloth
A detailed tool-calling benchmark covering various quantizations of the Qwen3.6-35B-A3B model has been published in the LocalLLaMA community on Reddit. The evaluation consisted of 144 test runs on a V100 GPU cluster (8 GGUF variants × 3 KV cache quantizations × 2 context lengths × 3 repetitions).
The main results are as follows:
ByteShape GPU-5** (4.15bpw) posted the highest overall score, outperforming the larger Unsloth UD-Q4_K_M and UD-Q6_K
q8_0 KV cache quantization is a nearly free lunch** — the difference from f16 is within the margin of error
Long contexts caused an average 10-point score drop across all quantizations, though ByteShape GPU-5 proved relatively resilient
Meanwhile, a PR has been submitted to llama.cpp adding MTP (Multi-Token Prediction) support for the smaller Gemma 4 variants (E2B/E4B), raising hopes for faster inference on mobile and edge devices.
---