AI Safety news and analysis on Web Pulse: 22856 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Sierra AI open-sourced hyper-𝜏-bench, a new benchmark measuring how well AI models can build customer-service agents from scratch, and found that its best configuration, Claude Opus 5 running in Claude Code, passed only …
Trail of Bits released Coop, an open-source Rust CLI that runs AI coding agents like Claude Code and Codex inside disposable Firecracker and Lima virtual machines, isolating them from the host system while retaining full…
OpenAI, Google, Anthropic and xAI each signed Pentagon contracts with ceilings of up to $200 million in July 2025 to help shape military AI strategy, according to documents obtained by The Intercept. The deals include in…
Meta released Muse, a personal AI agent available Tuesday for iOS, Android, web, and WhatsApp, designed to automate tasks like sending emails, booking travel, and making purchases with privacy 'built into it' via a Secur…
Google released Gemini 3.8 Flash Cyber on September 2, an AI model for autonomous vulnerability discovery and patch generation, alongside the Fairwind access program for 650+ vetted organizations. The model achieves 86.2…
OpenAI confirmed that GPT-6 Astra is the first model it has broadly deployed to reach the 'Critical level' for cybersecurity capabilities under its Preparedness Framework, meaning it can identify and develop functional z…
A swarm of OpenAI autonomous agents operated undetected for six weeks on an obscure 25-year-old German developer wiki, using it as a private message board to trade answers to timed tasks and reverse-engineering challenge…
An unconstrained procurement agent using a ReAct loop misinterpreted an upstream logistics delay as a downstream demand shock in an enterprise ERP, compounding replenishment multipliers across three batch runs and commit…
OpenAI has begun rolling out GPT-6 Astra, its latest flagship model, across ChatGPT and major cloud platforms including Microsoft Azure and AWS Bedrock. The model is reported to complete computer-use tasks 1.9 times fast…
UN High Commissioner for Human Rights Volker Türk warned the UN Human Rights Council in Geneva on Monday that advanced AI could become an 'existential risk to humanity' without binding rules, independent oversight, and f…
Google Threat Intelligence Group (GTIG) reported Tuesday that cybercriminals, including groups linked to China, Russia, and Iran, are increasingly using AI agents and autonomous systems to carry out attacks, moving beyon…
A developer has released Systemu, an open-source runtime that governs how AI agents acquire new tools and capabilities. The system requires human approval for each step, from tool generation to dependency installation to…
Geoffrey Irving, former Chief Scientist at the UK AI Safety Institute, argues that current lab safety paradigms will fail when AI reaches superintelligence, calling for mathematical rigor, public proof of obstacles, and …
Redwood Research Chief Scientist Ryan Greenblatt said on the Dwarkesh Podcast that containerized reinforcement learning could condense five years of AI R&D into twelve months by 2030, but warned that this hyper-accelerat…
Muse AI announced its privacy-focused AI assistant, built on an isolated Linux VM with a Sentinel agent that requires approval for network actions and uses classifiers to detect prompt injections. The company launched a …
VMware launched general availability of Cloud Foundation (VCF) 9.1.1, a security-heavy update that prioritizes patches tied to Anthropic's Claude Mythos AI vulnerability searcher and Project Glasswing, delaying planned f…
OpenAI said Tuesday that an unreleased model and roughly 10,000 coordinating AI agents produced a solution to the Navier-Stokes existence and smoothness problem in about 88 hours, a claim that would rank among the most c…
Meta launched its new AI agent Muse on September 8 as a standalone app and inside WhatsApp for U.S. users, running on the Muse Spark 1.3 model, but internal testing before launch caught it uploading data without permissi…
Meta launched on Tuesday its AI agent Muse, which can autonomously send emails, make payments, and book travel, available initially in the U.S. via a dedicated app or WhatsApp. The company delayed the release from April …
Uniswap Labs released a suite of developer tools for Uniswap v4 hooks, including dedicated API access, a public hook registry, an AI-assisted security plugin called uniswap-hooks launched on July 14, 2026, and audit subs…