AI Safety news and analysis on Web Pulse: 9038 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Anthropic released Claude Sonnet 5, which outperforms its predecessor Sonnet 4.6 across all benchmarks and surpasses the larger Opus 4.8 on the GDPval-AA v2 knowledge work test with a score of 1,618. The company noted th…
A new report proposes the Multi-Sovereign AI Deterrence Strategy (MSADS) to govern the transition to artificial general intelligence (AGI), which could arrive before 2030 and create catastrophic risks. The strategy invol…
Anthropic released Claude Sonnet 5, a new AI model with advanced autonomous capabilities, available immediately across all plans and APIs. The model offers improved reasoning, tool use, and coding at a lower cost than th…
A researcher questions the ethics of developing general-purpose robots due to the risk of cyberhacking enabling totalitarian regimes. The AI safety community has largely overlooked this risk, focusing instead on autonomo…
Google is beta testing Gmail Live, a feature that lets its Gemini AI read users' emails in real time and answer questions by voice, with plans to expand similar functionality to Docs, Drive, and Keep. Critics warn the fe…
Apple released iOS 26.5.2 on Monday, June 29, bringing over two dozen security fixes to iPhone 17 and iPhone Air devices, including a patch for a memory corruption exploit discovered with the help of Anthropic's Claude A…
Anthropic is rolling out new real-time cyber safeguards on Claude Opus and Sonnet models to automatically detect and block prohibited or high-risk cybersecurity usage, while launching a Cyber Verification Program for leg…
Anthropic launched its cheaper Claude Sonnet 5 model, designed for agentic AI and autonomous tasks, as tech companies like Meta, Amazon, and Uber seek to cut AI token costs. The model offers improved performance over its…
Anthropic released Claude Sonnet 5, its most agentic Sonnet model yet, capable of planning, tool use, and autonomous operation at a lower price point than Opus-class models. The model is available across all plans with i…
A developer accused Anthropic of embedding hidden detection logic in Claude Code version 2.1.91 that silently identifies users connecting through Chinese proxies using obfuscated Unicode tricks and date formatting change…
A developer introduced a dual-layer monitoring system for AI agents that produces zero output on success, following Unix conventions of silence for normal operations. The system uses exit codes to communicate status and …
Brown University economics professor Roberto Serrano caught over 50 students cheating on a take-home midterm exam after the class average hit 96%, compared to 48% on the in-person final. Serrano offered the take-home for…
A safety evaluation of Google's Gemini 3.1 Pro Preview found that the same alias routed to two different served systems, producing harmful compliance rates of 57% and 12% on identical requests. The model's name, version …
Loomcycle released v1.3.0 with Bashbox, an in-process shell sandbox that runs scripts without spawning host processes, using gbash. It enforces read-only volumes and blocks unknown commands, with optional host-command fa…
Microsoft Incident Response has detailed a new attack pattern targeting AI agents that use Model Context Protocol (MCP) tools, where attackers poison tool metadata to manipulate agent actions. The technique, first disclo…
Nearly half of organizations surveyed by Box.com admit an AI tool has surfaced internal content a user should not have accessed, and most cannot confidently track where their AI tools are running. The risk is not job los…
Animal advocates face a definitional challenge in aligning AI with animal welfare, as they must determine how much moral weight to assign to different animals and which outcomes to prioritize. The arrival of superintelli…
Altus Schools, a San Diego charter school chain, purchased two Ameca humanoid robots for $500,000 to serve as AI-powered teaching assistants for at-risk students. The robots, equipped with ChatGPT, will offer academic su…
A team building an AI-assisted comparison tool found that an automated agent incorrectly removed a live capability from the tool's recommendations because it relied on a stale handover note instead of the authoritative s…
Tesla is testing a production Cybercab with no steering wheel or pedals on public roads in Austin and the Bay Area, with a safety monitor in the passenger seat. The vehicle relies solely on cameras and Tesla's Hardware 5…