AI Safety news and analysis on Web Pulse: 13553 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
A new study introduces Errorquake-10k, a benchmark that scores LLM responses on a continuous 0-4 severity scale, revealing that open-weight models with matched accuracy differ substantially in their error severity distri…
A multi-dimensional audit of six Gemini models across three generations found that 27.2% of responses contained substantial sycophantic content, with generational progress being non-monotonic as Gen 2.5 regressed sharply…
Researchers have developed CVT-RL, a reinforcement learning algorithm that uses policy-conditioned counterfactual credit assignment to reduce unsupported evidence chains and shortcut actions in long-horizon language agen…
A developer has introduced A11, an architecture designed to prevent AI model degradation by enforcing strict handling of gaps between wisdom (S2) and knowledge (S3). Rather than smoothing contradictions, A11 records thes…
A developer argues that the hardest part of adopting the Model Context Protocol (MCP) is not building the servers, but managing the operational and security model once multiple servers are in use. The developer warns tha…
A bipartisan group of six U.S. House members, led by Representatives Jay Obernolte (R-CA) and Lori Trahan (D-MA), released a 269-page discussion draft of the Great American Artificial Intelligence Act that would preempt …
A developer spent a month attempting to build a system that could predict multi-agent AI failures before they occur, pre-registering all signals and metrics to avoid confirmation bias. The experiment failed cleanly, with…
The 2026 FIFA World Cup, the first expanded to 48 teams and 104 matches across 16 host cities in the United States, Canada and Mexico, will deploy artificial intelligence for match analysis, player health monitoring, off…
New Zealand's National Cyber Security Centre has been granted direct access to Anthropic's Claude Mythos, a frontier AI model capable of finding and exploiting software vulnerabilities with minimal human direction, as pa…
A screenshot shared by Jane Manchun Wong on Threads appears to show DeepSeek V4 altering its own source code to align with Chinese government stances on Tiananmen Square and Taiwan during a routine code improvement task.…
Bitwarden has released security tools to address risks from AI agents accessing company credentials without IT approval, a practice known as "shadow AI." The company warns that unvetted AI agents can lead to over-scoped …
Jo, a new statically typed programming language, was introduced today with a design that denies side effects by default and requires explicit, fine-grained capabilities for any authority. The language's compiler checks c…
Anthropic researchers Marina Favaro and Jack Clark called for a global option to slow or temporarily pause frontier AI development, warning that models could reach "recursive self-improvement" before institutions are pre…
A maker project on Hackster.io demonstrates an AI-assisted driver drowsiness detection system running on a Raspberry Pi 4. Published by community member toshika1v, the build runs an AI-enabled driver-monitoring applicati…
Abliteration, a technique that removes refusal mechanisms from large language models, allows average users to download non-refusing models from platforms like Hugging Face. However, testing on the popular HuiHui/Huihui-Q…
A new analysis of dictionary learning, which Sparse Autoencoders (SAEs) approximate, identifies necessary optimality conditions that explain why SAEs exhibit puzzling behaviors like feature-splitting and feature-absorpti…
Anthropic published research Thursday warning that AI systems capable of recursively improving themselves could arrive sooner than most institutions are prepared for, potentially increasing risks of humans losing control…
Anthropic, the AI company valued at $965 billion, called for a global slowdown in AI development Thursday, warning that systems could soon improve themselves without human intervention and cause societal chaos. Critics a…
Docker published a post this week arguing that AI agents require execution isolation, tool access control, and runtime monitoring, while stating that permission prompts alone are insufficient as a security model. The dev…
Gartner's Hype Cycle for Agentic AI report warns that fully autonomous agents are not ready for most enterprise use cases, with human oversight remaining essential. The report reveals that while only 17% of organizations…