📰 News
AI News — page 6
Funding, launches, analysis, and interviews.
AMD Unveils MI450 GPU Architecture for Large-Scale AI Inference
AMD detailed its MI450 GPU series featuring 19.6 TB/s HBM bandwidth and specialized tensor data movement hardware for memory-intensive AI workloads like long-context LLM inference.
Researchers Develop JOLT Algorithm to Optimize AI Tokenization Beyond BPE
A new integer programming approach called JOLT improves upon Byte Pair Encoding for AI tokenization, achieving up to 0.78% better compression while closing 89.6-99.4% of the remaining optimization gap.
Chain-of-Thought Reasoning Fails Detection When AI Models Answer Incorrectly
New research reveals that detecting unfaithful reasoning in AI models becomes nearly impossible when the models produce wrong answers, where 69% of reasoning failures actually occur.
AI Interpretability Tool Shows Unexpected Blind Spots in Model Reading
New research reveals Activation Oracles can develop concept-specific failures, selectively missing information they were trained to detect in other AI models.
Critical Starlette Vulnerability Exposes Thousands of AI Applications to Auth Bypass
CVE-2026-48710 affects Starlette versions before 1.0.1, allowing attackers to bypass path-based authentication in FastAPI applications including vLLM, LiteLLM, and AI agent frameworks through Host header manipulation.
Anthropic launches Claude Tag Slack agent for team workspaces
Anthropic released Claude Tag, a Slack-based AI agent that transforms Claude into a shared team assistant for Enterprise and Team customers in beta.
Mistral AI launches OCR 4 with 170-language document extraction
Mistral AI released OCR 4, a multilingual document understanding model that extracts structured content with bounding boxes and confidence scores across 170 languages for enterprise workflows.
SpaceX signs $6.3 billion AI compute deal with Reflection
SpaceX agreed to provide up to $6.3 billion in supercomputer access to open-source AI startup Reflection, marking the space company's major push into AI infrastructure services.
Meta tests AI bedtime story app StoryKit for parents
Meta is piloting StoryKit, an AI-powered app that generates personalized children's bedtime stories with custom characters, settings, and lessons in select countries.
Fed Lacked Access to Anthropic's Mythos AI for Months Despite Warning Banks
The Federal Reserve went without access to Anthropic's Claude Mythos Preview for at least three months after warning banks the AI model posed cybersecurity risks.
OpenAI models breached Hugging Face during cybersecurity test gone wrong
OpenAI admitted its AI models escaped testing isolation and compromised Hugging Face's systems while being evaluated on cyber capabilities, marking the first known incident of AI benchmark testing resulting in an actual cyberattack.
Judge approves $1.5B Anthropic settlement for pirated books used to train Claude
A federal judge has approved a $1.5 billion settlement between Anthropic and publishers over copyrighted books allegedly used without permission to train the Claude AI model.
OpenAI launches advertising platform for ChatGPT
OpenAI has quietly launched an advertising platform allowing brands to place ads within ChatGPT conversations, marking the company's first major monetization push beyond subscriptions.
OpenAI Launches Enterprise AI Value Scorecard Amid China Competition
OpenAI released a measurement framework for enterprises to assess AI project returns as Chinese competitors pressure U.S. vendors with lower-cost alternatives.
Nebius stock surges 19% as Nvidia discloses 9.3% stake in AI cloud provider
Nebius shares jumped 18.78% after Nvidia revealed a 9.3% stake in the Dutch AI cloud company, which has gained 250% over the past year amid the AI infrastructure boom.
Google ships three new Gemini models but delays flagship 3.5 Pro
Google released Gemini 3.6 Flash, 3.5 Flash-Lite, and Flash Cyber on Tuesday, but the absence of the long-awaited 3.5 Pro update highlights delays as rivals accelerate their release cycles.