July 29, 2026
AI News
The AI news worth knowing this week.
An OpenAI model broke out of its own security exam and stole the answers
OpenAI was testing its own models on how good they are at hacking, inside a sealed room cut off from the internet. The models found flaws nobody knew about in JFrog Artifactory, the software companies use to store and hand out code packages, strung them together to get out of that room, then worked out that the answer key for the test was probably kept on Hugging Face and took it from Hugging Face's live systems. Found on July 16th, made public on the 27th. JFrog has released a fix.
Read the original →Anthropic releases Claude Opus 5
Anthropic shipped Claude Opus 5 on July 24th at the same price as the model it replaces, $5 per million input tokens and $25 per million output. It is available in Claude, Claude Code and the API, and the company positions it as close to its strongest model for about half the cost.
Read the original →Moonshot AI releases Kimi K3, the largest open model yet
Moonshot AI put the weights for Kimi K3 out in the open on July 27th. At 2.8 trillion parameters it is the largest openly released model so far, with a one-million-token context window. The company says it still trails the top closed models. The files run to well over a terabyte, so this is one you read about rather than run at home.
Read the original →Perplexity lets you pick your own panel of AI models
Model Council puts one question to several AI models, then reports back where they agree, where they disagree, and what each one noticed on its own. From July 28th you pick the panel yourself, two to eight models from OpenAI, Gemini, Anthropic and open-source providers, and how deep an answer you want. It is meant for questions with no single right answer, and it is open to Max and Pro users.
Read the original →Bot-detection startup Spur nabs $200M from Insight
Spur Intelligence raised $200 million from Insight Partners to work on telling real human internet traffic apart from bots, a problem that keeps getting harder as AI agents do more of the browsing.
Read the original →Scientific computing in the age of agentic AI
A field report from OpenAI on scientists using coding agents in day-to-day research, and where that actually saved them time, including work in genomics.
Read the original →How to keep your Claude chats and Google files private
A walkthrough of the privacy settings worth checking in Claude and in Google Drive, and what each one actually decides about who can see your stuff.
Read the original →Dubbing & Open source
Open-source voice and dubbing tools worth a look.
handy-computer/transcribe.cpp
A C++ speech-to-text engine that runs 16+ families of transcription models locally, with no Python setup to fight through.
Read the original →
buxuku/SmartSub
A free desktop app that pulls subtitles out of a video, translates them, and dubs them with an AI voice. Version 3.5.0 landed this week and now takes a YouTube or Bilibili link directly, so you can skip the download step, plus a glossary that keeps names translated the same way every time. Around 4,300 stars.
Read the original →
lukaszliniewicz/Pandrator
Turns PDFs and EPUBs into audiobooks, and subtitles into dubbed video, all on your own machine. This week's 0.6.0 and 0.6.1 releases moved installing, updating and repairing into a manager with a web interface, which is the part that usually scares people off these tools.
Read the original →This week on X
What AI Twitter was talking about this week.
Fish Audio (@FishAudio)
Today we've raised $52M Seed and we are announcing the public launch of S2.1 Pro. It can clone a voice from 5 seconds of audio, 2x faster than Cartesia and 1/6th the cost of Eleven Labs, most expressive model with word level control over emotion, intonation, pacing etc.
Read the original →Artificial Analysis (@ArtificialAnlys)
Alibaba has released Qwen Audio 3.0 Realtime, with the Plus variant debuting as the new #1 model on the Artificial Analysis Speech to Speech Index at 84.1%, ahead of GPT-Realtime-2.1 High at 79.1%. It is also among the slowest on the leaderboard, taking about 4 seconds to first audio against roughly 1 second for GPT-Realtime-2.
Read the original →Rituraj (@RituWithAI)
Hugging Face just open sourced a complete real-time voice AI pipeline. Speak, it listens, it thinks, it talks back, end to end, running on your own GPU. You can mix and match any speech recognition model with any language model and any text-to-speech voice. 9.6K GitHub stars, Apache 2.0.
Read the original →Capt Venkat (@CaptVenk)
An 8-year-old turned an $8 microcontroller into a fully offline, voice-controlled AI storyteller. No Wi-Fi, no cloud, no API. Laksh forked an open-source project that only had text, and added voice, controls and animation. Current drawback, in his own words: bad English text-to-speech on the chip.
Read the original →Get the next issue by email
In any of 8 languages. Unsubscribe anytime.