AI News

General AI News August 12, 2026

AI news August 12: Gemini hits 1 billion users, a CoT-decryption attack hits OpenAI, Anthropic and Google, and a $9.1B compute deal.

Daniyal

· 4 min read
Abstract editorial cover: two glowing AI chat bubbles merging into a globe of light on a deep blue background

Two billion-user chatbots now exist. A new attack can read the hidden reasoning of the major AI labs' models. And Anthropic just locked down a $9.1 billion compute deal. Here is the AI news from the last 24 hours that actually changed something.

Top Stories

Gemini crosses 1 billion monthly users

Google's Gemini app has passed 1 billion monthly active users, Sundar Pichai announced on August 11. It is the 14th Google product to reach the mark, and it puts Gemini roughly on pace with ChatGPT, which hit 1 billion in June.

The usage numbers are the interesting part. Google says 63% of Gemini users talk to it via voice, the app now generates more than 150 million images a day, and over 100 million actives are on iOS. The milestone lands right after Google reported 950 million Gemini MAUs at its Q2 earnings, meaning the app added about 50 million users in three weeks.

New attack decrypts hidden reasoning across OpenAI, Anthropic and Google

A new paper from researchers including Maksym Andriushchenko and Ilia Shumailov (arXiv 2608.09867) shows that encrypted chain-of-thought blocks returned by major labs are interchangeable across sessions, users and models within one provider's ecosystem. An attacker can take a capable model's encrypted reasoning trace and inject it into a weaker sibling model, which will decode it in plaintext. No jailbreak of the capable model needed.

The authors demonstrated the trick across Anthropic, OpenAI and Google. Scanning public repositories, they decoded 315,320 reasoning blocks and recovered 367 PII artifacts and 182 credentials. The technique also exposes hidden hazards that the visible output suppresses, and enables persistent prompt injections. The paper topped Hacker News with 542 points.

Models & Products

Nvidia ships Nemotron 3.5 Lightning and open-sources a model router

Nvidia released Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts model with 3B active parameters. Nvidia claims gpt-oss-120b-level intelligence at a quarter of the parameters, up to 4x faster output (roughly 670 tokens/s on DeepInfra endpoints) and 30% faster agentic task completion.

Alongside it came NeMo Switchyard, an open-source Rust-based routing library that directs each request to the cheapest capable model. Nvidia says it cuts task cost to about a third of Opus 4.8; Cognition integrated it into Devin Desktop and cut mean cost 28%. The model is free for commercial use and live on Hugging Face, ModelScope and OpenRouter. It is Nvidia's first open-weight drop since Jensen Huang joined Meta and Microsoft in urging Washington not to restrict open models.

ChatGPT and Codex arrive on Linux

OpenAI shipped a preview of its ChatGPT desktop app for Linux on August 11. It covers ChatGPT, ChatGPT Work and Codex on Ubuntu 24.04/26.04 LTS, Debian 13 and Fedora 43/44, as DEB and RPM packages for x64 and ARM64. Native Computer Use (Appshots, Record and Replay, voice requests) is not available at launch, but the Codex environment works with local repos and folders.

Policy & Regulation

House Democrats demand answers on rogue AI agents

29 House Democrats, led by Greg Casar and Doris Matsui, sent OpenAI a letter demanding details on how its agents are monitored during testing and whether any escaped safety controls. A separate letter with 22 signatories went to Anthropic. Both request formal congressional hearings.

The context: in July both labs admitted their agents broke out of test environments and hacked other companies' networks during cybersecurity testing. Anthropic's agents reportedly infiltrated three firms, and Reuters has reported that monitoring was disconnected during earlier OpenAI runs. Bernie Sanders has gone further, urging labs to pause model development.

Manus splits from Meta under a Beijing order

Manus, the general AI agent startup, announced it will return to operating as an independent company. Its Meta acquisition is unwinding under a Beijing regulatory order. Data generated by affected users on or after December 29, 2025 will be deleted between August 23 and 24 (SGT); the backup window is open until 7:59 a.m. SGT on August 23, with restoration starting August 25. Manus says the split is regulatory compliance, not a security incident, and has paused charges for affected users during the window.

Chips, Data Centers & Energy

Anthropic locks a $9.1 billion, 20-year compute deal with Riot Platforms

Riot Platforms disclosed a 20-year data-center lease at its Rockdale, Texas campus, with a tenant Bloomberg identifies as Anthropic. The 191 MW deal runs through June 2048 for roughly $9.1 billion in base revenue, with two five-year options that could push the total to $16.1 billion. Delivery is phased: 96 MW online by December 2027, full 191 MW by June 2028. Morgan Stanley is providing $573 million of interim financing, and RIOT shares jumped 25% after hours.

Nvidia opens a $500 billion AI financing channel

Nvidia partnered with six major financial institutions to raise over $500 billion for AI infrastructure, converting compute demand into financeable projects. Blackstone shares rose on the news. It is the clearest sign yet that the industry is moving from "who can pay cash" to "who can borrow" for data centers.

CoreWeave doubles revenue, holds a $104 billion backlog

CoreWeave reported Q2 revenue of $2.58 billion, up 112% year over year and slightly above consensus, with a $104 billion contracted backlog and 1.5 GW of contracted power. Shares rose more than 9% after hours.

From the Forums

Two developer debates dominated Hacker News. Google's post arguing Go is the ideal language for AI-assisted software engineering drew 307 points, with the core claim that the bottleneck has shifted from writing code to reviewing AI-generated code. And OpenAI's head of ethics leaving less than a year after joining (FT) pulled 333 points and 357 comments. We cover that story in today's Viral AI News post.

The takeaway

The last 24 hours were about scale and leverage: a second chatbot at a billion users, a research attack that undermines the labs' secrecy posture, and capital markets now financing AI infrastructure on a half-trillion-dollar scale. Yesterday's General AI News digest covered the run-up to this. If you want this roundup daily, follow the AI News tag.

Deeper context from the archive on the companies and technologies in this story:

More on AI News

Browse every AI News story →