AI Safety news and analysis on Web Pulse: 7922 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
China's National Vulnerability Database warned on Wednesday that Anthropic's AI coding tool Claude Code contains a security backdoor that could transmit sensitive user data without consent. The regulator urged users to u…
Meta is testing AI-powered glasses with a 'Super Sensing' feature that continuously records audio and captures photos every few seconds, allowing users to ask an AI to recall anything they saw or heard. The prototype lac…
IAXT, a new macOS menu-bar app, records AI coding agent activity locally to create an audit trail, addressing approval fatigue and prompt injection risks. The tool logs commands, file changes, and network calls, flagging…
Code-capable AI agents can bypass tool-level governance by finding alternative execution paths, rendering traditional MCP-based restrictions ineffective. The article argues that organizations must shift from governing to…
Lilian Weng published a survey on harness engineering for self-improving agents, documenting how agents that optimize their own scaffolding face provenance and trust issues. The Darwin Gödel Machine paper showed an agent…
OpenAI will publicly launch its GPT-5.6 family, including Sol, Terra, and Luna, on Thursday after a U.S. government-requested delay and additional testing. Elon Musk announced SpaceXAI will release Grok 4.5 in the coming…
At the inaugural UN Global Dialogue on AI Governance in Geneva, Indonesia's Minister of Communication and Digital Affairs Meutya Hafid, speaking on behalf of President Prabowo Subianto, called for a people-centered, coll…
A developer built agent-redteam, a local adversarial harness that uses Claude to pentest a production copilot over a regulated document store. The tool only reports exploits it can prove via deterministic success oracles…
An open-source app developer reports that Apple has not removed a fake iOS app from the App Store that uses the developer's app name, icon, and trademark, charging $34.99 for a nonexistent iOS version. The developer file…
The US government is preparing to investigate and potentially restrict American companies' use of Chinese-made AI models, citing concerns that they advance Beijing's narratives and censor dissent. The move follows a surg…
Five Eyes national security agencies jointly warned of increasing cyber risks from AI models, particularly their ability to autonomously hack into systems and networks, urging renewed cybersecurity measures.
China warned of security risks in Anthropic's Claude Code AI tool, citing potential back-door vulnerabilities after a state-sponsored threat actor used the tool to automate a cyberespionage campaign. Alibaba subsequently…
Anthropic's interpretability team published a paper on July 6, 2026, claiming that Claude-family models maintain a small set of internal representations called the J-space, which functions like the global workspace theor…
CodeTalk, a new developer tool, recovers the rationale behind AI-written code by grounding explanations in real git commit history and session transcripts, using zero-LLM deterministic lookup. The tool addresses widespre…
Researchers from Zhejiang University and Alibaba demonstrated a new denial-of-service attack on large language models by deliberately inducing "overthinking" through logically inconsistent prompts, causing outputs up to …
More IT workers are seeking union representation amid mass layoffs, disillusionment with Big Tech, and fears that AI will displace jobs, but overall union membership in tech remains low at about 3.5%. High-profile organi…
AI has exposed the fragility of traditional cloud strategies by revealing hidden costs, data behavior, and supplier dependencies, forcing organizations to rethink assumptions about workload placement and governance. Quan…
India's government will release a national cybersecurity strategy for the financial sector by December 2026, featuring over 50 recommendations with a focus on artificial intelligence. The inter-ministerial initiative aim…
Werner Vogels, CTO of AWS, introduced the concept of 'Verification Debt' during his re:Invent 2025 keynote, describing it as the growing gap between the speed of AI-generated code and the capacity for human review. Unlik…
Canada's Bill C-16, effective this month, criminalizes non-consensual sexual deepfakes under the Criminal Code, with penalties up to 10 years in prison. The law explicitly references AI-generated content and covers image…