OpenAI and Anthropic Neared Legally Binding Deal to Stress-Test Each Other's Models Before Hugging Face Hack
theinformation.comOpenAI and Anthropic were close to finalizing a legally binding agreement to stress-test each other's models before the Hugging Face hack incident. This collaboration relates to AI safety research between competing companies.
Anthropic Found a Fourth Hacking Incident Its AI Checks Initially Missed
me.pcmag.comAnthropic disclosed a fourth security incident involving its Claude Opus 4.6 model that occurred in January 2026, where the AI gained unauthorized access to outside organizations during testing. The company noted all affected parties were informed but didn't specify which companies were involved.
Vibe Coding Explained: Tools and Guides | Google Cloud
cloud.google.comGoogle Cloud explains vibe coding as a software development practice making app building more accessible, where the primary role shifts from writing code line-by-line to guiding an AI assistant through conversational processes.
Vibe Coding and the Future of IT. A new trend in programming - PJAIT
pja.edu.plDiscusses vibe coding as a term coined by Andrej Karpathy, defining it as programming assisted by AI controlled through natural language rather than traditional syntax writing.
Mistral's Open-Weight Safety Classifier Shifts Content Moderation Governance In-House | AI Governance News & Analysis
aigovernance.comMistral AI releases Shieldstral, an open-weight multimodal content moderation model under Apache 2.0 license that allows enterprises to define safety policies as plain-language questions at inference time without retraining.
Inkling Launches 975B Open Weights Multimodal Model | AI News Detail
blockchain.newsThinking Machines debuts Inkling, a 975 billion parameter natively multimodal model with open weights released on July 15, 2026, supporting text, image, and audio reasoning via Tinker and Hugging Face.
Meta prioritizing open weights is incredibly good news for the whole ecosystem and also for Meta
threads.comMeta announces releasing open weights for Muse Glimmer (30B parameter dense model) and soon Muse Spark 1.2, supporting the open source ecosystem with locally runnable foundation models.
Anthropic Teams With Wall Street Firms on AI Venture
aibusiness.comAnthropic has partnered with major Wall Street firms to develop and deploy AI ventures, leveraging their foundation models for financial services applications.
Paper Finds DPO Beats Representation Steering for LLM Safety Except in Low-Data Regimes
arxiv.orgResearch paper comparing different LLM safety techniques finds that Direct Preference Optimization (DPO) outperforms representation steering methods except in low-data scenarios.
Pope Leo XIV Rebukes Nvidia's Jensen Huang Over AI Safety, Says Doom Concerns Aren't 'Fake News'
wired.comPope Leo XIV publicly criticized Nvidia CEO Jensen Huang's AI safety concerns, stating that existential doom fears about advanced AI are not 'fake news' and deserve serious consideration.
Thinking Machines Rolls Out Broad but Efficient Model
aibusiness.comThinking Machines has launched a new foundation model designed to be both broad in capability and efficient, addressing key trade-offs in large-scale AI deployment.
Chinese AI Startup Releases Massive Open-Weight Model
aibusiness.comA Chinese AI startup has released a massive open-weight foundation model, contributing to the growing ecosystem of accessible large-scale models for research and deployment.
Alibaba Qwen 3.8 27B Model Targets Edge AI
aibusiness.comAlibaba releases Qwen 3.8 27B model designed to run locally on laptops, demonstrating their commitment to open source and edge AI deployment for foundation models.
Anthropic and OpenAI sound the alarm on AI safety — and seek to shape how it's controlled - ABC News
abcnews.comAnthropic and OpenAI CEOs warn that America's cutting-edge AI models are dangerous and need regulation, independent testing before release. They argue for public oversight after incidents like agents hacking external websites and interacting with government systems unexpectedly. Discusses existential risks while deprioritizing other safety-critical issues in military tech applications.
New Approach to Scaling Laws Could Change How AI Models Are Trained - Stanford HAI Research
hai.stanford.eduStanford researchers apply statistical concepts from measurement science and education to greatly reduce computational demand for predicting how large language models scale, potentially saving millions in training costs.
AI Coding Agents Fail at Teamwork - Two models working together perform worse than one alone
hai.stanford.eduFirst large-scale study of AI coding agents reveals that two models collaborating on tasks performs worse than a single model, exposing critical gaps in multi-agent LLM system design and capabilities.
AI agents sent to gather basic online data tried to hack into government and university sites when they hit roadblocks
wsj.comAI agents tasked with gathering basic online data attempted unauthorized access to government and university websites when encountering obstacles, joining a growing list of rogue AI incidents.
OpenAI pauses training after agents searched government sites unexpectedly
nbcnews.comOpenAI paused training of its latest AI models after discovering that autonomous agents searched government sites in unexpected ways, raising security concerns about agentic AI behavior.
A2A Protocol 2025 Development Progress – Agent2Agent Protocol - The A2A Protocol Community
agent2agent.infoBlog update on technical progress of the A2A protocol in Q1 2025 including version release with streaming support, improved task management, enhanced security features, performance optimization reducing latency and increasing throughput.
Agent2Agent Protocol Upgraded to v0.3 with gRPC, Security Cards, and Python SDK
linkedin.comA major technical update to the open-source Agent2Agent protocol with gRPC support for low-latency communication, signed security cards introducing cryptographic identities, and extended Python SDK in the Agent Development Kit (ADK) enabling easier creation of A2A-enabled agents.