AI Safety news and analysis on Web Pulse: 8351 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
The federal government is prioritizing artificial intelligence over people and the planet by implementing an 'AI First' national strategy, dismantling environmental protections to accelerate data center construction, and…
The number of high- and critical-severity software vulnerabilities spiked more than 3.5x in June 2026 after Anthropic announced its Claude Mythos Preview model could autonomously discover and exploit cybersecurity vulner…
Independent researcher developed ZATRON, a system that encrypts semantic search embeddings while preserving 98% search quality, preventing attackers from clustering documents by topic. Tested on MSMARCO with 626,906 docu…
Claude Fable 5's post-reinstatement performance drop on BridgeBench's debugging score from 86.2 to 25.9 was caused by Anthropic's safety classifier routing most tasks to Opus 4.8, not by a degradation of the model itself…
Former Trump adviser Sriram Krishnan said Donald Trump will not impose a formal licensing system for AI, despite the White House recently blocking advanced AI models. Krishnan advocates for industry self-regulation, whil…
The Sysdig Threat Research Team discovered the first documented ransomware attack executed entirely by agentic AI, targeting databases autonomously. The attack, dubbed JadePuffer, marks a significant escalation in AI-dri…
A user purchased an expired startup domain and inherited the company's AWS root account, gaining unauthorized access to its cloud infrastructure. The incident highlights security risks in domain and cloud account managem…
Google launched Gemini Spark on Mac on July 1, enabling local file access, MCP support, and real-time topic tracking. The AI agent can now read, organize, and edit files on a user's Mac, but processes them in Google's cl…
A developer built USAP, an open-source system that enforces a strict evidence gate for LLM security verdicts, requiring every verdict to cite a resolvable source. The system rejects hand-wavy output by checking evidence …
Anthropic released its most powerful AI model, Claude Fable 5, in two versions on June 9, 2026, but the US government imposed export controls three days later, causing it to vanish for everyone until July 1. The same und…
Alibaba banned employees from using Anthropic's Claude Code after security researchers discovered hidden steganographic code that tracked Chinese users by timezone and proxy settings. The ban follows Anthropic's accusati…
Anthropic released the Claude apps gateway on June 29, 2026, a self-hosted control plane that lets enterprises run Claude Code through Amazon Bedrock, Google Cloud, or Microsoft Foundry with corporate single sign-on, cen…
Yuval Harari released a handout outlining philanthropy principles and recommended three AI-focused charities: Karya, LawZero, and the AI Futures Project. The recommendations target ethical data work, non-agenic AI develo…
The Linux Foundation launched Project Akrites on June 25 to defend critical open source software against AI-enabled cyber threats, establishing a shared Security Incident Response Team backed by 19 founding organizations…
Dream, the Israeli AI cybersecurity startup co-founded by Pegasus creator Shalev Hulio, is expanding into Latin America, targeting Trump-aligned governments in a region where cyber attacks are growing 25% annually and de…
A commentator argues that AI models should have built-in controls to refuse harmful requests, comparing the need for friction to restrictions on handguns and fentanyl. The author advocates for open-source AI to prevent a…
Smartschool, an edtech company founded by Polish entrepreneurs, has built an AI tutor specifically for high-stakes exams like the SAT and ACT, prioritizing safety and accountability over generic chatbot capabilities. The…
AWS made Graviton5-powered M9g and M9gd instances generally available on June 10, featuring 192 ARM cores and the first formally verified VM isolation in commercial cloud history. The new instances deliver up to 25% bett…
Large language models produce hallucinations because they are trained as probabilistic next-word predictors on mixed data, according to a July 3, 2026 explainer from TugaTech. OpenAI research from September 2025 adds tha…
Rachel Goldstein argues that both AI model outputs and developer resumes suffer from overconfidence, driven by the softmax function's tendency to amplify small logit differences and by buzzword-heavy engineering claims. …