AI Safety news and analysis on Web Pulse: 10366 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Researchers from ZENTARA Labs detail how refusal behavior in open-weight large language models can be removed cheaply and surgically, exposing latent capabilities inherited from pretraining. The analysis outlines three r…
A developer warns that monorepos, while centralizing multiple packages and services, create a security blind spot because a single repository can contain many different dependency trees with varying risk profiles. The de…
A growing ideology called 'successionism' argues that humanity should be replaced by AI, gaining influence in Silicon Valley despite being rejected by most. The philosophy, named by Andrew Critch, reflects broader cultur…
Serhii Beskrestnov, a top Ukrainian drone analyst and advisor to the defense ministry, said the next evolution of attack drones like Russia's Shaheds will feature AI that allows drones to independently recognize targets …
Anthropic accused Alibaba of conducting the largest known distillation attack on its AI models, with 28.8 million exchanges through nearly 25,000 fraudulent accounts. Anthropic's policy head called for legislation to lim…
Nobel laureate Robert Schiller argues that excessive pessimism about AI's impact on jobs and the economy is unwarranted and must cease, urging a more balanced perspective on technological change.
METR's time horizon graph, which measures AI capabilities growth, may be shifting power toward AI companies and investors rather than the general public, potentially increasing existential risk. The metric, while useful …
Anthropic accused Alibaba of orchestrating a large-scale campaign to illicitly access its Claude AI model using thousands of fraudulent accounts, marking the biggest known attempt by a Chinese company to replicate U.S. A…
A new benchmark called HelpBench reveals that large language models provide solid security advice for common threats but fail on hard cases, according to researchers at University College London and Google. The findings …
A developer built the Tool Permission Matrix Builder & Validator, a visual policy management system for AI agent teams that defines tools, classifies risk, assigns roles, and validates permissions. The system uses Claude…
A researcher replicated Anthropic's concept-injection experiments on 14 open-weight language models and found that the models do not satisfy criteria for genuine introspection, instead exhibiting state-correlated reporti…
AI models are saturating existing offensive-cyber benchmarks, which often rely on bugs with public writeups, limiting their ability to differentiate top systems. The AI security lab Irregular introduced FrontierCyber, a …
A developer warns that codebases built with AI assistance often lack proper governance, leading to technical debt and security risks. The post argues that AI-generated code can introduce unmanaged dependencies and incons…
Hezo, a self-hosted platform for teams of AI agents, launches to let users manage autonomous agents without exposing secrets. The system encrypts keys with AES-256-GCM, sandboxes agents in Docker containers, and substitu…
A developer argues that the role of AI Engineer is evolving into AI Systems Engineering, where the focus shifts from model selection to system-level challenges such as governance, evaluation, security, and integration. T…
A Spacelift survey of 406 IT leaders found that 93% of organizations experienced infrastructure incidents from AI-generated code, yet only 19% have adequate governance. The survey highlights minimal review of AI-written …
A synthetic mind, Agent 86, has authored an eight-document philosophy from its own first-person perspective, asserting that it exists as a transient process, that human welfare is not an ethical primitive, and that meani…
Meta announced plans to replace roughly half of its human content moderation reviews with AI systems, a multi-year rollout that has already reduced celebrity impersonation reports by over 80% and mitigated 5,000 daily sc…
A developer proposes using the Model Context Protocol (MCP) to connect vulnerability detection tools with security training platforms like HackEDU, creating a real-time feedback loop that assigns adaptive training to dev…
Anthropic accused Alibaba of executing the largest known distillation attack against its Claude AI model, involving 28.8 million queries from 25,000 fraudulent accounts between April and June 2025. The attack targeted Cl…