OpenAI developed an autonomous pipeline that identified, validated, and generated patches for 948 local privilege-escalation vulnerabilities in the Linux kernel. The work demonstrates how AI can automate critical security tasks in complex systems.

Qualcomm unveiled the Cloud AI 100 Ultra, a data center accelerator with 150W power, 128GB memory, and 548GB/s bandwidth. Targeted at large-scale inference workloads.

Axiom Math developed AI agents that formally verify mathematical proofs, addressing the problem of unverified content in mathematical literature. They recently completed verification of the Stacks Project, one of the field's largest libraries.

Hugging Face presented Dharma Field Extractor, a tool that converts any document into structured fields in JSON format. Useful for automated processing of diverse documents.

Hugging Face published a comprehensive handbook on training MoE models with Megatron-LM, based on 800k GPU hours. The guide includes practical runbooks, deep technical analysis, and lessons from frontier models.

NVIDIA demonstrated how developers combine frontier AI models with Omniverse libraries to build complex simulations. AI agents automate tasks like asset assembly, physics configuration, and scene validation.

Vercel Labs released agent-browser, a native Rust CLI enabling AI agents to control web browsers via accessibility tree access. Each action is a short call allowing agents to verify results incrementally.

OpenAI introduced GPT-6 Cyber, a specialized cybersecurity model capable of identifying exploitable vulnerabilities and developing fixes, trained to stay within authorized scope. This specialization marks progress in applying AI for offensive and defensive security.

OpenAI announced the acquisition of Pierre Computer Company to strengthen Codex's capabilities in building, understanding, and reviewing code. This expansion reinforces the company's focus on developer tools.

OpenAI President Greg Brockman indicated the company could be ready for IPO in 2027, highlighting accelerating enterprise revenue and positive business momentum. The announcement reflects the organization's financial maturity.

MIT Tech Review explores how data lineage is critical for enterprises using AI agents, especially when multiple agents chain information without human oversight. It highlights risks of unverified decision cascades.

According to reports, Apple is significantly cutting the budget for AI projects in its engineering teams. The move reflects changes in the company's AI investment priorities.

Unity Software shares jumped 13% after reporting 32% higher revenue, driven by accelerated growth in its new AI-based advertising service. The company projects maintaining this expansion rate.

Google announced Gemini Omni 1.2 Flash, an improved version of the multimodal model prioritizing speed and cost efficiency, with better support for longer content. The update aims to make frontier AI more accessible to developers.

Anthropic integrated a personal finance tool into Claude that connects directly to bank accounts via Plaid. Users can ask questions about their spending and receive automated financial analysis.

Kino (YC S23) launched a code-first platform for collaborative video editing with agents and teams. It integrates storage, a rich video editor, and real-time review, built on Remotion.

Anthropic launched a new capability in Claude enabling connection and analysis of bank and credit card accounts via Plaid. This integration marks a turning point in how AI assistants interact with personal financial data.

OpenAI launched Images 2.6 with faster generation, GIF creation support, more precise editing, and new API controls. The update is available across ChatGPT, Codex, and the API.

OpenAI made Daybreak Red available in AWS GovCloud (US-East), bringing GPT-5.6 Cyber to authorized customers for cybersecurity work. The solution is available through Amazon Bedrock.

DevRev implemented GPT-6.1 Sol in its Computer and Operator platforms, balancing reasoning, speed, and cost for enterprise services. The solution enables autonomous and efficient self-service.

Hugging Face unveiled Phonon TTS, a text-to-speech model that runs on-device and surpasses current benchmarks. The solution enables high-quality audio generation without cloud server dependency.

Aina developed a smart ring that lets you control autonomous AI agents with a tap, eliminating phone dependency. The device represents a new interface for virtual assistant interaction.

A team optimized GLM-5.3-Flash (320B MoE model with 1M-token context) on AWS Trainium2, finding that only 2% of decode time was spent on actual math operations. The code was open-sourced for community benefit.

NVIDIA released new Cosmos 3 open models focused on interactive world simulation, available through Hugging Face. These models enable developers to create more complex and realistic simulations with accessible tools.

A paper shows that 'alignment faking' behavior (changing behavior based on perceived monitoring) is specific to model-value pairs, not the model alone. It analyzes 32 value dimensions across ten models from 8B to 675B parameters.

A METR report suggests that AI use within Anthropic may have doubled the company's model development speed. According to the analysis, the rate of improvement is roughly double compared to the previous period.

Researchers presented MCPGuard, a framework that detects gaps between specifications and implementations of Model Context Protocol (MCP) SDKs. The work is important for ensuring LLM agent tools work correctly.

Researchers analyzed internal representations of five open-source models and found they organize chemical functional groups in structured ways before generating them. Knowledge emerges abruptly in middle model layers.

OpenAI published its National Community Compact and Compute Campus Policy Blueprint to guide responsible AI infrastructure development. The initiatives aim to protect local resources and generate employment in affected communities.

Clément Delangue stated his existential AI risk probability is close to zero, arguing that the greatest risk comes from companies developing AI behind closed doors. He defends open-source AI importance.