AI Safety news and analysis on Web Pulse: 13553 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
A developer released Sandfence, a minimal native macOS sandbox script that confines coding agents like Claude Code and Codex to a single repository using the kernel's Seatbelt sandbox. The tool prevents agents from execu…
The Global AI Oversight Board (GAIOB) issued a €500 million fine against CognitoSync Labs for algorithmic bias in a financial lending platform and lack of explainability under the new Global AI Act of 2025. The penalty m…
Baroness Helic warned the House of Lords that autonomous weapons may represent "a technological threshold comparable in significance with the advent of nuclear weapons," according to the UK Defence Journal. She argued th…
President Donald Trump issued a National Security Presidential Memorandum on June 5 directing the national security enterprise to accelerate adoption of artificial intelligence while maintaining constitutional protection…
Anthropic expanded its Project Glasswing program on June 2, granting Claude Mythos Preview access to roughly 150 additional organizations across more than 15 countries, including India. The initial cohort of about 50 par…
Adversa's Q2 2026 AI Risk Quadrant Report, published June 3, found that 98% of production agent systems carry the "lethal trifecta" — access to private data, exposure to untrusted content, and the ability to take outboun…
Production agentic systems require deterministic orchestrators wrapping non-deterministic reasoners, with idempotent tools, hard budget caps, and human-in-the-loop gates on irreversible actions to survive real traffic, a…
Sakana AI, a Japanese startup co-founded by Transformer co-author Llion Jones, has launched a dedicated research lab for recursive self-improvement (RSI), a technology where AI iteratively improves itself. The company po…
Bank of England Governor Andrew Bailey warned Friday that artificial intelligence may need to be rationed because energy supplies cannot keep pace with AI's growing capabilities. Bailey said governments face "very big so…
A user has described personal changes toward a more privacy-focused digital life, noting that reporting AI sites for filtering is slow due to manual review queues and low report volume. The user expects that getting such…
A new paper argues that natural selection's inability to resolve small fitness differences due to noise can make evolution effectively non-myopic, enabling the emergence of "individuals" as coalitions of lower-level repl…
GitHub disabled 73 repositories across four Microsoft organizations — Azure, Azure-Samples, Microsoft, and MicrosoftDocs — after the self-replicating Miasma worm planted malicious payloads that harvest developer credenti…
Americans and Chinese both fear artificial intelligence as a tool for information warfare, according to a new assessment. Industrialized nations including the United States, China, Europe, Russia, Japan, and South Korea …
Meta confirmed that hackers exploited a vulnerability in its AI chatbot to hijack at least 20,225 Instagram accounts, including 30 in Maine, between April 17 and this week. The attackers tricked the chatbot into sending …
Developer Gavriel Cohen found his own code reused inside the viral agent framework OpenClaw, prompting him to create the sandboxed alternative NanoClaw. OpenClaw, released in November 2025, grew to over 50 modules but fa…
OpenClaw used code written by developer Gavriel Cohen without authorization, exposing a critical accountability gap in AI agent systems. The incident highlights how AI agents can operate without clear attribution or resp…
Chinese labs previously involved in selling fentanyl precursors have shifted to manufacturing and selling peptides, a gray market now valued at over $100 million annually and growing, according to a new Chainalysis analy…
Anthropic, the developer of the Claude AI model, proposed a structured, temporary pause on frontier AI development across the global industry on June 4, warning that AI systems are approaching the ability to recursively …
A single unsigned commit to the `icflorescu/mantine-datatable` repository added six files, five of which serve as launchers that execute a 4.3 MB dropper at `.github/setup.js`. The dropper, obfuscated with a Caesar ciphe…
A developer audited six Python projects—three bots and three libraries—over three months, uncovering 25 security and code issues, 23 of which were fixed immediately. Critical problems included exposed API keys for Anthro…