Microsoft shut down dozens of GitHub code repositories for Azure and AI coding tools after a reported hack that targeted AI developers' passwords.

对六个月内针对检测API的对抗性输入分析显示，最具破坏性的提示注入攻击主要针对通用训练（general-purpose training），而非对齐训练（alignment training）。有效的攻击利用模型的「助手属性」（helpfulness）、对话连贯性（conversational coherence）以及善意互动（good-faith engagement）作为攻击向量。

Apple is waiving cloud API costs for developers with fewer than 2 million first-time App Store downloads to make AI experimentation more accessible.

Apple announced a revamped Apple Intelligence platform architecture built on foundation models co-developed with Google using Gemini technologies. The models are designed to run on-device and via Apple's Private Cloud Compute infrastructure.

Xiaomi's MiMo-v2.5-Pro-UltraSpeed model is introduced as a 1 trillion parameter model capable of generating 1000 tokens per second. The article URL and Hacker News discussion link are provided.

According to a report from runtimewire.com, DeepSeek V4 Pro demonstrates superior precision compared to GPT-5.5 Pro in a benchmark test. The article includes a link to the original source and a Hacker News discussion thread.

OpenAI 近期限制用户下载聊天记录功能，即使账户已使用强密码和基于OTP的双因素认证（2FA）保护，仍需提供手机号才能导出。用户反映，未提供手机号将无法完成聊天记录导出操作。

The article discusses containing Claude across products, with a link to the full article at https://www.anthropic.com/engineering/how-we-contain-claude and a comments section at https://news.ycombinator.com/item?id=48392082.

Anthropic has published an open-source framework designed to use AI for discovering vulnerabilities in code. The project is available on GitHub, with community discussion hosted on Hacker News.

此安排出现在AI实验室与五角大楼围绕Claude模型的法律争端期间