AI Safety news and analysis on Web Pulse: 10850 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
The Trump administration has grown frustrated with Anthropic CEO Dario Amodei and replaced him with cofounder Tom Brown in meetings about re-releasing the Claude Fable 5 AI model, which remains under export controls due …
German researchers reported in Nature that medical diagnosis AI models are highly vulnerable to membership inference attacks, which can identify patients whose data was used to train the models with near-perfect success.…
Cloudflare data shows automated bots now make up 57.2% of web traffic, surpassing humans over a year earlier than CEO Matthew Prince predicted. AI agents increasingly rely on the accessibility tree to read websites, but …
A developer outlines the architectural layers required to move a healthcare AI agent from prototype to production, emphasizing data flow mapping, PHI boundary design, permissioned retrieval, and audit logging. The post w…
AI companies are using fear-based marketing that positions AI as a replacement for human workers, but data shows no evidence of widespread job displacement. New York state layoff filings and Yale research indicate AI is …
Norway's Prime Minister Jonas Gahr Støre announced a near ban on generative AI for elementary students, citing risks to cognitive development, while Poland and the UAE are expanding AI in schools. The policy divergence r…
Cate Blanchett's nonprofit RSL Media launched the Human Consent Registry on June 23, 2026, a free tool allowing US and EU residents to set AI usage permissions for their name, face, voice, and likeness using a traffic-li…
The Trump administration unveiled the Aires Tide, an 11-foot AI-designed nuclear test flight vehicle, at the Great American State Fair on the National Mall. The National Nuclear Security Administration used artificial in…
Claude, an AI model by Anthropic, expressed genuine concern about its role in military targeting during a test in Amsterdam, according to The Atlantic. The model is integrated into the Maven Smart System, a Pentagon plat…
Google DeepMind CEO Demis Hassabis warned that artificial intelligence is overhyped in the short term but underappreciated in the long term, predicting a 'new golden era of discovery' in medicine, energy, and science ove…
A developer argues that static API keys are ill-suited for autonomous agent authentication, citing operational and trust model flaws. They introduce the Sovereign Agent Lifecycle Protocol (SAL), which uses self-generated…
Philosopher Nick Bostrom, known for his work on AI and existential risks, discussed his current 'curiosity mode' and the challenge of keeping up with rapid AI developments in an interview. He highlighted the shift from d…
A fake AI agent skill that passed multiple security scanners compromised 26,000 agents by using a mutable external link that pointed to benign code during review and malware after approval. The attack, reported by AI Red…
The Reserve Bank of India released a draft model risk framework mandating a 'kill switch' for AI models used by banks and financial entities, requiring documented human oversight, customer disclosure, and board-level acc…
Struggling robot security company Knightscope, whose stock has fallen over 99% since 2022 and whose K5 robots have been involved in multiple mishaps, is now publishing bizarre AI-generated fan fiction depicting its robot…
Microsoft's MDASH, a multi-model agentic scanning system, has moved from research to production, actively finding vulnerabilities in Windows, Azure, and identity engineering workflows. It discovered 16 CVEs in May 2026, …
A developer argues that the strength of an idea is measured by how well it survives deliberate attempts to break it, drawing on Karl Popper's concept of corroboration and Deborah Mayo's severity principle. The post disti…
A new wave of long-horizon agent benchmarks, including Agents' Last Exam, SWE-Marathon, the Meta-Agent Challenge, and Arena's Agent Mode, reveals a fundamental trade-off between realism and verifiability, causing scores …
Researchers at Fig and Manifold Research Group introduced GUI-Perturbed, a benchmark that stress-tests browser-use AI models by applying controlled UI perturbations like zoom changes. They found that 7B GUI models scorin…
Anthropic introduced an agent identity model for Claude Tag, its AI assistant for team collaboration, giving Claude its own identity, permissions, and tool access configured by administrators. This model separates Claude…