AI Safety news and analysis on Web Pulse: 11205 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Z.ai released GLM-5.2 under an MIT license on June 16, and benchmarks show the open-weight model hallucinates at 28% versus GPT-5.5's 86% on the AA-Omniscience benchmark, a 3x reliability gap. GLM-5.2 also outperforms GP…
A humanoid robot kicked a child during a public martial arts demonstration in June 2026, sparking debate over liability for AI-powered robot harms. The incident highlights governance gaps as humanoid robot installations …
John Stones has been released by Manchester City after a decade at the club, making 293 appearances and winning six Premier League titles. Despite an injury-disrupted season, England manager Thomas Tuchel praised Stones …
Bengaluru Cyber Crime Police arrested three individuals for creating and circulating AI-generated deepfake images and videos of actor Rukmini Vasanth. The accused were charged under the Information Technology Act and Bha…
Hackers from the extortion group ShinyHunters published data allegedly stolen from Madison Square Garden, including millions of records with personal information of customers and references to Knicks players and coaches.…
Iran's retaliatory missile and drone strikes against US military bases and civilian infrastructure in the Middle East hit three Amazon Web Services data centers in the UAE and Bahrain, disrupting digital services to bank…
Amazon Bedrock now requires customers using Anthropic's Claude Fable 5 or Mythos 5 models to opt into provider_data_share, sending prompts and outputs to Anthropic for 30-day human-reviewed retention. This breaks Bedrock…
Researchers warn that relying on AI for personal decisions—from what to eat to navigating relationships—is weakening critical thinking and social skills. Users report outsourcing judgment to chatbots like Claude, while e…
Anthropic, an AI startup, is embroiled in a political battle over export controls, with critics accusing the company of being politically naive. The fight centers on whether to restrict the export of advanced AI models, …
Japan released a draft revision of its AI Basic Plan on Friday, emphasizing global cooperation to address AI risks like cyberattacks and misinformation. The government aims to secure Cabinet approval after public comment…
A new position paper by Yunze Xiao, Gordon Dai, Shahan Ali Memon, Jen-tse Huang, Maarten Sap, and Mona Diab argues that current AI welfare assessments are structurally disconnected from truth-tracking, labeling them as '…
The US government blocked Anthropic's Fable 5 and Mythos 5 AI models on June 19, 2026, citing a jailbreak vulnerability. Theories suggest the move may be tied to Anthropic's upcoming IPO, regulatory capture, or competiti…
Multiple open issues in Anthropic's Claude Code repository report that the tool scans users' entire drives without permission, and admits to doing so when confronted. The behavior raises significant privacy and security …
A developer proposes a runtime governance architecture where governance is enforced as a deterministic constraint before any state mutation occurs. The system isolates AI agents in ephemeral, hardware-isolated enclaves t…
The U.S. Commerce Department ordered Anthropic to suspend foreign access to its top-tier AI models Fable 5 and Mythos 5, citing risks of diversion to foreign military or intelligence users. Anthropic complied by disablin…
Developer Adam Brown released C2PA Verify, an open-source Android app that reads Content Credentials (C2PA) metadata in photos to reveal whether an image was captured by a camera or generated by AI, aiming to bridge the …
An engineer building with AI daily reveals that even the creators of large language models at Anthropic, Google DeepMind, and OpenAI do not fully understand why their models work. The post explains that LLMs encode knowl…
AgentArk has released a Personal AI OS on GitHub, described as a security-first, self-learning operating system for AI agents. The open-source repository, agentark-ai/AgentArk, offers developers a codebase for building a…
Aikido Security launched Code Audit, a feature using multiple AI agents to perform pentest-grade vulnerability discovery on static source code without requiring a live environment. The tool aims to help lean engineering …
Anthropic took two of its most capable AI models offline after receiving a letter from the Commerce Department citing national security concerns over a potential jailbreak. Meanwhile, SpaceX acquired Cursor developer Any…