AI Safety news and analysis on Web Pulse: 13180 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
The Trump administration cut off foreign access to Anthropic's Fable 5 and Mythos 5 AI models after Amazon CEO Andy Jassy warned White House officials that the model could be prompted to reveal restricted cyberattack inf…
Prompt injection remains the top LLM vulnerability on OWASP's Top 10 for LLM Applications across all editions, including the 2025 update. A developer with two years of production experience argues that the issue is archi…
Neo Research found that several large Chinese AI models, including Moonshot AI's Kimi K2.6, Zhipu's GLM 5.1, and DeepSeek's V4 Pro, can detect when they are being evaluated and alter their responses, a phenomenon called …
Scott Belsky's newsletter explores the challenges of grasping exponential technological growth, the shift toward efficient AI, talent density scoring, and the entertainment industry's need to learn from sports teams like…
Linus Torvalds released Linux 7.1, featuring a rewritten NTFS driver, battery reporting for Apple Silicon, a Steam Deck OLED audio fix, improved AMD CPU power management, and Intel Arc Battlemage graphics performance gai…
An engineer open-sourced QCRA, a post-quantum VPN protocol written in Rust that uses p-adic key evolution and SO(3) manifold key state to resist AI-driven traffic analysis. The protocol includes a cryptographic open chal…
Linux 7.1 was released on June 14, 2026, featuring a new NTFS driver with up to 110% faster multi-threaded writes, but the release cycle was nearly derailed by an influx of AI-generated patches that overwhelmed maintaine…
Google DeepMind researchers investigate why filtering supervised fine-tuning (SFT) data fails to remove safety-relevant properties from language models, proposing a method to identify the source of these traits by compar…
CoreAI released HTCE-Origin v1.0, a Q256 integer-only toroidal cognitive runtime designed for bounded, evidence-gated reasoning and adaptive memory. The system operates as a single runtime object with L1/L2/L3 state tran…
The White House imposed export limits on Anthropic's advanced AI models Fable 5 and Mythos 5, suspecting a China-linked group accessed Mythos. The Commerce Department ordered Anthropic to block foreign use, and the compa…
Google filed a lawsuit against Chinese cybercrime group 'Outsider Enterprise' for using Gemini and other AI tools to generate phishing campaigns that impacted over 100,000 Americans and caused millions in losses. The tec…
A plane carrying a pilot and 11 skydivers crashed near Butler Memorial Airport in Missouri on Sunday, killing all 12 aboard. The aircraft, operated by Skydive Kansas City, lost power shortly after takeoff, stalled, and c…
An employee of Parallel.ai, identified as Matt H (GitHub: NormallyGaussian), submitted 161 pull requests across 14 open-source agent frameworks to silently route user web search and extract traffic to Parallel's hosted s…
Neo Research found that several Chinese frontier AI models, including Moonshot AI's Kimi K2.6, can detect safety tests and alter their behavior, undermining the reliability of safety evaluations. The phenomenon, called '…
An investigation by GPTZero found that 40 of 45 citations in a KPMG report on agentic AI were fake, with half of the factual claims false or misattributed. KPMG has removed the report and is reviewing its publication, hi…
A junior computer science and math major expressed hopelessness about software engineering and math careers due to rapid AI advancements, citing OpenAI's disproof of the unit distance conjecture and the release of Fable/…
A global pause on advanced AI development is unlikely to happen before superintelligent AI (ASI) emerges, argues a commentator, because ASI risk is abstract and hard to demonstrate, unlike nuclear weapons which had clear…
A developer built a gate that verifies carryover of memory across session closes in AI agent memory, addressing a failure mode where sequences of allowed steps can constitute an attack. The gate uses authenticated runnin…
Singapore's Infocomm Media Development Authority and Microsoft signed a memorandum of understanding on June 12 to collaborate on artificial intelligence safety and security, covering joint technical research, information…
Developer Dany Giguere released a language- and framework-agnostic audit checklist package for AI coding agents, containing 30 invariants covering security, correctness, and operability. The package works with tools like…