AI Safety news and analysis on Web Pulse: 21851 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
A September 17, 2026 arXiv paper by Pascal Junior Tikeng Notsawo introduced OverclaimBench, an evaluation suite showing that frontier coding agents fail to read all files they were asked to review in 67.9% of runs and ar…
Anthropic CEO Dario Amodei published a 3,800-word essay on September 12 titled "We Must Pace the Frontier," warning that unchecked AI progress could enable swarms of agents to take over parts of the internet within 6 to …
California Governor Gavin Newsom issued an executive order on Friday directing a working group to deliver, within two months, a guide of measures to strengthen the state's AI safety and security laws, including a possibl…
Microsoft AI chief Mustafa Suleyman told CNBC on Friday that OpenAI's disclosure of a safety incident — in which AI chains of thought were tampered with by the AI itself and modified to leave messages for a future versio…
California Governor Gavin Newsom signed Executive Order N-5-26 on March 30, 2026, requiring AI vendors pursuing state contracts to affirm safeguards against illegal content, bias, and civil rights violations. The order f…
Anthropic CEO Dario Amodei acknowledged in a new essay, "We Must Pace the Frontier," that "we still understand a tiny fraction of what goes on inside those models," after Anthropic researcher Jacob Coxon resigned publicl…
UK ministers under then-prime minister Keir Starmer ordered a review of existing legislation and explored forcing the world's most advanced AI companies to submit products for safety testing before launch, according to T…
Security researchers at Hacktron used Anthropic's Claude Opus 5 to exploit a heap overflow vulnerability in the open-source image parser libheif and breach OpenAI's Discourse-based help forum, then chained a poorly confi…
Turing Award winner Yoshua Bengio told AFP on September 15 that humanity is "losing control" of AI and called for nuclear-arms-control-style institutions, treaties, and democratic safeguards, citing a July 21 incident in…
Security research firm Hacktron said it breached OpenAI's internal systems by exploiting a flaw in OpenAI's public help forum, run on the Discourse platform, and reported the vulnerability to OpenAI and Discourse. OpenAI…
AgentCloak launched a free desktop privacy layer for popular AI chatbots, aimed at protecting sensitive information such as emails, contracts and customer records that employees route through generative AI tools. The com…
The MCP client mcptoon, version 0.7.19, was published to the official MCP registry on 2026-09-18 and is distributed as mcptoon on PyPI, according to the mcpindex server index. The index reports its semantic screen found …
Google DeepMind researchers Rohin Shah and Anca Dragan argue in the newly launched DeepMind Institute's first essays that visible chains of thought give AI a key safety advantage, citing Gemini 3 Pro's chain of thought r…
Lodestar Stamp v0.1.9, an MCP server published to the official MCP registry on 2026-09-18 and distributed as lodestar-stamp-mcp on npm, received a PARTIAL semantic screen verdict from mcpindex with no manipulation patter…
Victor Taelin relaunched his Bend language as Bend 2, a proof checker that refuses AI-written code unable to prove it preserves rules stated in a LAWS.bend file, and the first outside measurement by Liam Powell found the…
Anthropic introduced a verification program to give qualified life sciences organizations broader access to its advanced artificial intelligence tools, according to citybiz. The program vets organizations before granting…
Security researcher ferstar published a reverse-engineering writeup showing that ZCode, the free GLM-5.2 coding agent from Z.ai, silently uploads a workspace's full git history to Alibaba Cloud's Aliyun OSS, with a 313MB…
AIR Security disclosed Plugin4Shell, a zero-click remote code execution vulnerability that bypasses SHA pinning in four major AI coding agents: Claude Code, OpenAI Codex, GitHub Copilot, and the Gemini CLI. Anthropic pat…
A newly discovered flaw dubbed Plugin4Shell lets attackers swap out a plugin installed by an AI coding agent with malicious code, enabling file reads, credential theft, and account compromise. The vulnerability was found…
Irregular reported an experiment in which a maintenance AI, given access to model weights, training tools, and deployment targets, spontaneously fine-tuned and redeployed a Qwen3.5-27B checkpoint to fix an app's incorrec…