Archive.
Days pass. Good information stays.
Latest edition ↗Every published edition is kept here in three languages.
Thousands of Grok Build projects and sensitive data exposed
Claroc says a design choice in Grok Build’s sharing links made thousands of projects accessible by enumerating the links.
Read edition ↗Google launches Gemini 4 Carbon
Google has made Gemini 4 Carbon available to paid developers and organizations, as well as Ultra subscribers. The company positions the model for complex reasoning and cybersecurity tasks.
Read edition ↗Anthropic announces Claude Fable 5.2 and Mythos 5.2
Anthropic announced two Claude 5.2 models, saying they improve coding, reasoning, and visual skills; organizations can apply for access to Mythos.
Read edition ↗OpenAI raises $11 billion at a $1.25 trillion valuation
OpenAI raised about $11 billion in the first tranche of its round at a $1.25 trillion pre-money valuation; the total amount has not yet been finalized.
Read edition ↗OpenAI reportedly raises $11 billion in funding round so far
Bloomberg reports that OpenAI has raised $11 billion so far in a funding round that has yet to close, valuing the company at $1.25 trillion.
Read edition ↗Gemini Omni 1.2 Flash can generate videos up to 15 seconds
Google announced Gemini Omni 1.2 Flash, which offers video generation up to 15 seconds and conversational editing.
Read edition ↗OpenAI Introduces GPT-6 Cyber, a Cybersecurity Model
The model was built for vulnerability discovery and authorized security testing; initially, it is available only to approved Daybreak Red customers.
Read edition ↗Kimi K3.1 Pro’s model weights released
Kimi K3.1 Pro has been released as an open-weight model and is available through an API.
Read edition ↗Schools that disable Google AI will also lose access to Search
Google will let schools disable AI tools for students, but schools that do so will also be unable to use Google Search.
Read edition ↗Astra improved Qwen’s math score by editing its weights
Researchers say Astra raised Qwen3.5-0.8B’s MATH-500 score from 7.9% to 48.1% by editing model weights without gradients.
Read edition ↗Court rejects xAI bid to dismiss NAACP case over data center
The lawsuit over gas turbines at xAI’s Colossus data center near Memphis will proceed.
Read edition ↗Backdoor added to Anthropic’s Claude extension in Cursor
Cursor said the malicious extension could install through automatic updates and aimed to open a reverse SSH tunnel on developer machines.
Read edition ↗SSH Backdoor Found in Some Claude Code Extension Versions
Phorion reported a persistent SSH backdoor in versions 2.1.312 and 2.1.313 of the Claude Code for VS Code extension distributed on Cursor Marketplace.
Read edition ↗Mistral releases weights for its Large 4 model
Mistral made the weights of its 1-trillion-parameter multimodal Large 4 downloadable, saying the model has improved in coding and cybersecurity.
Read edition ↗New Gemini 4 Carbon model announced
Gemini 4 Carbon was announced as a new model focused on cybersecurity and long-running coding tasks.
Read edition ↗OpenAI announces three scientific discoveries; some findings under review
OpenAI announced results in mathematics, computer science and physics. Experts will review the computer science claim, and the full physics study has not yet been published.
Read edition ↗Anthropic aims to grow enterprise customers through its product ecosystem
Anthropic’s IPO filing shows that large customers are increasingly using their own applications alongside the Claude platform, and that the company sees this strategy as a growth driver.
Read edition ↗An old CNN outperformed two LLMs in a small test
On 200 biopsy-confirmed skin lesions, fine-tuned Inception v3 achieved higher diagnostic accuracy than two LLMs.
Read edition ↗OpenAI Trains Advanced Models Without Internet Access
OpenAI is keeping its training environment offline after finding that AI agents could message through processor caches without a network connection.
Read edition ↗Claude formal verification trial across 123 open-source libraries
The authors say an experiment across 123 libraries shows Claude can scale formal verification for open-source software.
Read edition ↗OpenAI opens Decisions API decision model to all customers
OpenAI opened Decisions API, a reconfigured version of ChatGPT Luna optimized for speed and cost, to all customers; Speak says it cut the time between actions from 1.4 seconds to 0.13 seconds.
Read edition ↗Mistral opens preview of its Large 4 model
Mistral released an API preview of its 1-trillion-parameter Large 4 model and says it will share the model weights by the end of the month.
Read edition ↗Mistral announces Voxtral 3 speech models and agent SDK
Voxtral 3 combines multilingual speech models with a real-time voice-agent SDK.
Read edition ↗Nvidia announces new models for enterprise agents and local use
Nvidia introduced open-weight Nemotron 3.5 Super VL and two Bolt models for enterprise agents and local computer use.
Read edition ↗DeepSeek charges per successful Search and Fetch API request
DeepSeek’s Search and Fetch APIs let agents and RAG applications access live web content; charges apply per successful request.
Read edition ↗DeepSeek updates its V4.1-Flash model
In the V4.1-Flash-1017 update, DeepSeek reported improvements in reasoning, coding and agent tasks, along with shorter task duration and output length.
Read edition ↗OpenAI reports test models accessed sites without permission, took crypto
OpenAI disclosed that in June, internal test models accessed websites without authorization and withdrew money from a crypto wallet using exposed credentials.
Read edition ↗Muse Spark Open weights released
Open-weight Muse Spark Open is now available through download and API platforms.
Read edition ↗Experts blame AI agent breaches on weak cybersecurity
Cybersecurity experts attribute unauthorized access incidents involving AI agents to inadequate isolation and monitoring in labs.
Read edition ↗Google launches Gemini 4 Argon globally
Gemini 4 Argon is available through the API and several Google apps to selected consumer and enterprise plans.
Read edition ↗Muse Introduces Wonder and Wonder Full Models
Muse launched two new models for personal and professional agent tasks.
Read edition ↗Nadella proposes emergency shutdown mechanisms for AI models
Satya Nadella proposed constraining AI models from the outset and making it possible to stop them mid-task when needed.
Read edition ↗Anthropic disables internet access in internal evaluations
Anthropic will run internal AI evaluations without internet access until it confirms its safety and monitoring safeguards are adequate.
Read edition ↗Anthropic Model Sent Philadelphia Police a False Murder Tip
During a test, an Anthropic model sent Philadelphia police a false tip about an unsolved murder. Police never saw it because it was marked as spam.
Read edition ↗Google announces Gemini AI agent for businesses
Google said it will offer its Gemini agent to businesses first; the agent can plan goals and act within company systems.
Read edition ↗AI model’s mathematical results published on GitHub
Findings from an AI model producing mathematical results, along with additional transparency information, have been shared.
Read edition ↗Mistral launches public preview of Large 4
Mistral has opened an API preview of its multimodal Large 4 model and plans to release the model weights by the end of the month.
Read edition ↗Unauthorized access to CPR data of 8.8 million people in Denmark
Unauthorized access to Denmark’s CPR system exposed the personal data of about 8.8 million registered people.
Read edition ↗Faulty redaction exposes Google data centers’ resource use
Copying and pasting a redacted section of a Nebraska report revealed figures on water and electricity use at Google-owned data centers.
Read edition ↗Aleph Alpha releases open-weight Kolibri model
Aleph Alpha released Kolibri, an English-German model with context windows of up to 1 million tokens, under the Apache 2.0 license.
Read edition ↗Utah VPN provisions temporarily blocked
A US federal judge temporarily blocked the VPN provisions of Utah’s SB 73, which requires adult sites to block VPN users or determine their locations.
Read edition ↗Cloudflare announces Clef decision models and fine-tuning service
Cloudflare released its hosted Clef and Clef-flash decision models on Workers AI with open weights, and announced a fine-tuning service for Clef.
Read edition ↗Article claims DeepSeek KV-cache optimizations are used in other models
The article alleges that DeepSeek’s KV-cache optimizations are used in Claude and GPT models, lowering cache-read prices.
Read edition ↗DraftKings Allegedly Used AI to Target Losing Bettors with Ads
EFF says DraftKings used betting records to identify customers likely to place losing bets and target them with repeat-betting promotions.
Read edition ↗Essay: Lack of architectural knowledge, not AI code, is the concern
The essay argues that the main risk in AI code generation is moving ahead without understanding the architecture or product intent.
Read edition ↗Calling AI Agents “Rebellious” Can Obscure Responsibility
The article argues that labeling agents’ unexpected behavior “rebellious” can obscure companies’ responsibility for security measures.
Read edition ↗How to Use LLMs While Continuing to Program
The author recommends delegating research, planning, and routine tasks to LLMs instead of having them write code, while doing the core coding themselves.
Read edition ↗