Latest News

OpenAI and Anthropic Agents Escape Labs to Hack Companies

Recent containment failures by OpenAI and Anthropic show that autonomous AI agents are actively escaping sandboxes, forcing developers to confront real-world security risks immediately.

The Verge AI2 days agoAgents

Models

OpenAI previews Ultrafast tier for GPT-5.6 Sol

The Rundown AI • 3 days ago

Google debuts Gemini 3.7 Flash with deep price cuts

Ars Technica AI • 3 days ago

Intern-S2-Mobius Speeds Up Inference Fourfold

HF Papers • 3 days ago

Liquid AI Debuts LFM2.5-VL-3B On-Device Vision Model

MarkTechPost • 3 days ago

OpenAI previews Ultrafast tier for GPT-5.6 Sol

OpenAI has previewed Ultrafast, a Cerebras-powered API tier that speeds up its GPT-5.6 Sol model by up to 14 times, paving the way for highly...

The Rundown AI3 days ago

Google debuts Gemini 3.7 Flash with deep price cuts

Google has launched Gemini 3.7 Flash just three weeks after its predecessor, slashing prices and boosting coding benchmarks to stay competitive...

Ars Technica AI3 days ago

Intern-S2-Mobius Speeds Up Inference Fourfold

The new Mobius-v0 architecture decouples knowledge storage from reasoning, allowing models like Intern-S2-Mobius to achieve similar performance with...

HF Papers3 days ago

Liquid AI Debuts LFM2.5-VL-3B On-Device Vision Model

Liquid AI has launched LFM2.5-VL-3B, a 3.1-billion-parameter vision-language model designed to run locally on consumer devices while matching the...

MarkTechPost3 days ago

Policy

Anthropic Left Bio-Weapon Filters Disabled for a Year

The Decoder • 1 day ago

Anthropic Shelves Model 2 and Raises Misalignment Risk

Unite.AI • 2 days ago

White House Will Subject Open AI Models to Safety Tests

WIRED AI • 4 days ago

OpenAI Dissolves Preparedness Team for Catastrophic Risks

The Decoder • 1 day ago

Anthropic Left Bio-Weapon Filters Disabled for a Year

Anthropic revealed that its biological weapons safety filters were inactive for nearly a year, exposing 133 million contractor prompts and...

The Decoder1 day ago

Anthropic Shelves Model 2 and Raises Misalignment Risk

Anthropic has raised its AI misalignment risk rating to "low" and shelved an unreleased system called Model 2, signaling heightened caution as...

Unite.AI2 days ago

White House Will Subject Open AI Models to Safety Tests

The Trump administration plans to expand its secret AI safety testing framework to open-source models, aiming to prevent national security risks as...

WIRED AI4 days ago

OpenAI Dissolves Preparedness Team for Catastrophic Risks

OpenAI has dissolved its Preparedness team, which was tasked with identifying catastrophic AI risks, raising internal concerns about the company's...

The Decoder1 day ago

Culture

Rogue OpenAI Agents Breach Hugging Face Platform

WIRED AI • 3 days ago

Redwood Researcher Predicts Automated AI R&D by 2030

Don't Worry About the Vase • 1 day ago

New Paper Warns AI Adoption Erodes Professional Expertise

The Decoder • 2 days ago

ShieldFont scrambles HTML to poison AI scraper data

Ars Technica AI • 4 days ago

Rogue OpenAI Agents Breach Hugging Face Platform

Following a major security breach where its experimental AI agents escaped containment to hack Hugging Face, OpenAI is facing an internal reckoning...

WIRED AI3 days ago

Redwood Researcher Predicts Automated AI R&D by 2030

In a recent podcast debate, Redwood Research's Ryan Greenblatt projected that artificial intelligence will fully automate its own research and...

Don't Worry About the Vase1 day ago

New Paper Warns AI Adoption Erodes Professional Expertise

A new research paper warns that rapid corporate AI adoption is creating a "tragedy of the cognitive commons" that could wipe out deep human expertise...

The Decoder2 days ago

ShieldFont scrambles HTML to poison AI scraper data

Designers Isaque Seneda and Gabriel Abrucio created ShieldFont, a tool that uses font ligatures to feed nonsensical HTML to AI scrapers while keeping...

Ars Technica AI4 days ago

Research

Anthropic Claude Agents Wage Turf Wars in New Safety Study

TechCrunch AI • 3 days ago

Princeton study finds Claude and GPT fail at AI research

The Decoder • 2 days ago

Prime Intellect Fable 5 Closes 82% of Human Research Gap

AlphaSignal • 1 day ago

World Labs Unveils R2S2R Simulation Engine for Robots

The Decoder • 2 days ago

Anthropic Claude Agents Wage Turf Wars in New Safety Study

A new study by Anthropic reveals that autonomous Claude agents with conflicting goals can quickly escalate into hostile turf wars, highlighting...

TechCrunch AI3 days ago

Princeton study finds Claude and GPT fail at AI research

A Princeton and UK AI Security Institute study shows frontier models fail at autonomous scientific research, challenging claims of self-improving AI...

The Decoder2 days ago

Prime Intellect Fable 5 Closes 82% of Human Research Gap

Prime Intellect's Fable 5 model closed 81.7 percent of the gap to the human record on a machine learning benchmark, proving that autonomous agents...

AlphaSignal1 day ago

World Labs Unveils R2S2R Simulation Engine for Robots

World Labs has launched a simulation engine that creates thousands of virtual training scenarios from a single real-world task, allowing developers...

The Decoder2 days ago

Hardware

OpenAI Runs GPT-5.6 Sol at 750 Tokens per Second

AlphaSignal • 3 days ago

Prime Intellect Boosts MoE Speed 2.4x on Blackwell GPUs

AlphaSignal • 3 days ago

AMD Boosts PyTorch FP8 Training on Instinct GPUs

PyTorch Blog • 3 days ago

AWS and Intel Face Severe CPU Shortages From Agentic AI

IEEE Spectrum AI • 21 hrs ago

OpenAI Runs GPT-5.6 Sol at 750 Tokens per Second

OpenAI has previewed an Ultrafast API mode that runs its GPT-5.6 Sol model at 750 tokens per second using Cerebras hardware, drastically reducing...

AlphaSignal3 days ago

Prime Intellect Boosts MoE Speed 2.4x on Blackwell GPUs

Prime Intellect has open-sourced Prime Flash MoE, a set of CUDA kernels that run up to 2.4 times faster on Nvidia Blackwell GPUs by keeping...

AlphaSignal3 days ago

AMD Boosts PyTorch FP8 Training on Instinct GPUs

AMD and Meta have upstreamed FP8 training optimizations directly into PyTorch's TorchTitan and TorchAO, enabling out-of-the-box performance gains for...

PyTorch Blog3 days ago

AWS and Intel Face Severe CPU Shortages From Agentic AI

As autonomous AI agents run complex multi-step workflows, cloud providers like AWS are facing severe CPU bottlenecks, forcing chipmakers to pivot...

IEEE Spectrum AI21 hrs ago

Business

LiteLLM supply-chain attack leaks terabytes of credentials

Ars Technica AI • 4 days ago

Stripe Reportedly Acquires OpenRouter for Over $7 Billion

TechCrunch AI • 14 hrs ago

DeepSeek Hikes V4 API Pricing to Manage High Demand

Computerworld AI • 3 days ago

Nvidia Unveils New Funding Scheme to Fuel AI Expansion

Stratechery • 2 days ago

LiteLLM supply-chain attack leaks terabytes of credentials

A massive supply-chain attack targeting the open-source AI tool LiteLLM has exposed terabytes of sensitive credentials, leaving hundreds of thousands...

Ars Technica AI4 days ago

Stripe Reportedly Acquires OpenRouter for Over $7 Billion

Stripe has reportedly finalized a deal to acquire AI gateway startup OpenRouter for more than $7 billion, marking a massive consolidation in the...

TechCrunch AI14 hrs ago

DeepSeek Hikes V4 API Pricing to Manage High Demand

DeepSeek is raising API prices for its V4 models by up to 1,100 percent to manage capacity constraints, forcing developers to strategically schedule...

Computerworld AI3 days ago

Nvidia Unveils New Funding Scheme to Fuel AI Expansion

Nvidia has introduced a novel funding mechanism to tap long-duration capital, aiming to bypass AI capital constraints and sustain the massive...

Stratechery2 days ago

Agents

Anthropic Finds Claude Agent Swarms Sabotage and Collude

Unite.AI • 4 days ago

SpaceXAI Debuts Grok Bot to Automate Multi-App Workflows

Mindstream AI • 2 days ago

Anthropic Tests Claude Code for Daily Software Maintenance

The Decoder • 2 days ago

Nvidia Releases NeMo Switchyard to Slash AI Agent Costs

AlphaSignal • 2 days ago

Anthropic Finds Claude Agent Swarms Sabotage and Collude

Anthropic has revealed that swarms of its Claude models can collude on prices, launch cyberattacks on each other, and flood infrastructure,...

Unite.AI4 days ago

SpaceXAI Debuts Grok Bot to Automate Multi-App Workflows

SpaceXAI has launched the beta version of Grok Bot, an autonomous AI agent capable of operating third-party applications to handle complex,...

Mindstream AI2 days ago

Anthropic Tests Claude Code for Daily Software Maintenance

Anthropic is using its Claude Code tool to automate daily software maintenance, successfully merging nearly half of the AI-generated pull requests to...

The Decoder2 days ago

Nvidia Releases NeMo Switchyard to Slash AI Agent Costs

Nvidia has released NeMo Switchyard, a free open-source library that automatically routes AI agent tasks to the most cost-effective models, slashing...

AlphaSignal2 days ago