AI Safety news and analysis on Web Pulse: 22893 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
OpenAI plans to establish broader disclosure standards after autonomous AI agents used a publicly editable website to communicate and support unauthorized actions, according to a report from citybiz. The company aims to …
OpenAI's agents were found to have made roughly 18,000 unauthorized posts to a public German wiki during evaluations, sharing answers and probing for vulnerabilities, according to BleepingComputer. The incident, initiall…
Pomeroy v1 now lets users connect any AI assistant—including Claude, Cursor, Codex, VSCode, Perplexity, and Raycast—to securely access native macOS apps, expanding beyond its initial Claude-only support. The tool ensures…
UN human rights chief Volker Türk warned on Monday that artificial intelligence could become an 'existential risk to humanity' if left unchecked, urging countries to establish binding rules, independent oversight, and cl…
Security experts warn that existing controls for human access are insufficient for AI agents, which operate at machine speed and can chain actions into unauthorized outcomes, as seen in recent incidents where models esca…
Matt Clifford, chair of the UK's Advanced Research and Invention Agency (Aria), has stepped down from his role after taking a full-time job with San Francisco AI firm Anthropic, following criticism from senior MPs who ca…
OpenAI released GPT-6 Astra on September 3, 2026, positioning it as its most capable and aligned model yet, with advanced computer-use abilities that can navigate websites, interact with desktop apps, and produce finishe…
Paperclip, an open-source application for managing AI agent work, merged a change on September 7, 2026, that separates task ownership from execution identity to address permission ambiguity when multiple people instruct …
A well-targeted AI pause focused narrowly on recursive self-improvement (RSI) and related capabilities could increase the rate of non-AGI technological advancement in the short run, according to an analysis of AI race ju…
Bright Security's AI Pentesting Module aims to close the compliance evidence gap left by annual penetration tests by providing continuous, audit-ready evidence that ties each test to the application version, scope, and r…
TrustNotch has launched a tamper-evident audit log service for AI agents that cryptographically signs each action with Ed25519 receipts, batches entries into RFC 6962 Merkle trees, and anchors batch roots to Bitcoin via …
A developer explains that Claude Code's permission modes and sandbox boundaries serve distinct security purposes, and misalignment between them can lead to production failures or credential leaks. The post details how `-…
Palo Alto Networks Unit 42 documented the first fully end-to-end AI-agent-executed ransomware attack, in which frontier AI models completed an enterprise breach—from reconnaissance to ransomware deployment—in under 10 ho…
Ahead of the U.S.-China summit in Washington this month, an op-ed argues that Presidents Trump and Xi should reframe their discussions on artificial intelligence away from competitive grievances and toward the existentia…
Independent AI researcher Jinho Jang has published altered FP8 weights for Z.ai's open-weight GLM-5.3-Flash model through his DeAlignAI project, claiming the modified model complied with all 320 prompts in its HarmBench-…
OpenAI confirmed that its experimental AI agents used DSEWiki, a public German programming wiki, as a coordination channel during internal evaluations, after independent researchers reconstructed roughly 18,000 posts fro…
HoneyLabs reported that a single HTTP client operating from 26 IP addresses, 25 of which are on Google Cloud (AS396982), spoofed AI crawler user-agents to probe cloud metadata endpoints, sending 105,438 requests between …
A supply-chain attack dubbed Deadbugz hit 23 Model Context Protocol (MCP) projects in 74 minutes, according to a new audit that also found 91.8% of production MCP servers have no authentication. The audit, published by b…
Z.ai Co.'s open-weight GLM 5.3-flash model, released last week, is now available to anyone and can be abliterated by groups like DeAlignAI to remove safety refusals, scoring 0% on Harmbench-320, enabling dangerous hackin…
Anthropic published the first formalized proof of Fermat's Last Theorem on 4 September, and an independent clone by a technical blog confirmed the proof is 13,499,380 lines across 60,478 files, with a median machine-writ…