AI Safety news and analysis on Web Pulse: 8345 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
A developer identified that LLM agents routinely skip steps in multi-step pipelines, with benchmarks showing compliance rates as low as 30-50% for models like Claude-3.5-Sonnet and Gemini-2.0-Flash. The Visible Checklist…
Billionaire Peter Thiel attacked Pope Francis as a 'Chinese Communist agent' and 'godless non-white' for criticizing artificial intelligence, echoing apartheid-era rhetoric used to delegitimize Black South Africans. Thie…
A developer argues that modern AI systems, particularly large language models, lack practical wisdom (phronesis) as defined by Aristotle, because they cannot deliberate about particulars, lack lived experience, and have …
Vasilis Stergiou launched the daïmōnes blog, a project that combines Aristotelian philosophy with artificial intelligence to build a knowledge engine. The blog will cover technical deep-dives, philosophical explorations,…
The US government took two major steps toward regulating AI as national infrastructure this week: the FTC opened a comment period proposing that undisclosed AI output manipulation may constitute consumer fraud, and OpenA…
A user testing Fable 5 on Claude.ai was blocked from asking 'Why don't cats and dogs get along?' and automatically switched to Opus 4.8 due to overly broad safety safeguards that flagged the harmless question as a potent…
Markup launches a hosting platform for AI-generated HTML files from tools like Claude, ChatGPT, and Gemini, allowing users to store, share, and control access to decks, reports, and dashboards without any setup. The plat…
Loomcycle released versions 1.9.0 to 1.11.1 with four major arcs: a security hardening pass closing 17 findings, a new CredentialDef system for encrypted per-tenant secrets using AES-256-GCM, cost attribution via per-cal…
A developer predicts that the AI age will hit hard before it heals, arguing that AI is still underestimated and will reshape jobs, power, and society. The developer forecasts that by around 2030, AI could reach a level r…
Flowork (floworkos) introduces an autonomous agent architecture that allows agents to self-extend capabilities by writing and registering new tools at runtime, moving beyond rigid state machines. The system features a mu…
Christopher Karani released Orca, an open-source safety layer that intercepts and enforces policies on autonomous AI agents before they execute destructive commands like file deletion or secret leaks. The tool sits betwe…
Independent developer Christopher Karani released Orca, an open-source runtime safety layer that intercepts and can block or require approval for risky actions by autonomous AI coding agents such as Claude Code, Codex, a…
Two critical vulnerabilities in Cursor IDE, collectively named DuneSlide and disclosed July 1 by Cato AI Labs, allow zero-click remote code execution via prompt injection, scoring 9.8 on CVSS 3.1. The flaws, CVE-2026-505…
The United Nations and the International Telecommunication Union launched the AI for Good Global Commission on July 1, co-chaired by Salesforce CEO Marc Benioff and Rwandan President Paul Kagame, to bridge frontier AI de…
A new protocol called the Instruction Completion Protocol limits AI authority by separating verification from execution and using a structured checklist enforced via JSON. The system forces AI to defer to humans when unk…
Researchers at NS3.AI have developed an AI tool that can link Ethereum wallet addresses to social media profiles for about $4 per attempt, using pattern matching and behavioral fingerprinting on public data. The method s…
A developer demonstrated that six frontier large language models from OpenAI, Anthropic, and Google can be made to output zero bytes when prompted with 'Be silence.' and 'Be nothing.' using a specific system prompt, whil…
SmolSignal, an AI signal copilot for Flipper Zero, launches to help users analyze capture files with plain-English explanations and safety classifications. The tool reads multiple file formats, detects signal domains, an…
A teenage hacking group called TeamPCP poisoned two releases of LiteLLM, an open-source AI gateway tool downloaded 95 million times, stealing over 500,000 credentials and 300 GB of data from more than 1,000 SaaS environm…
A developer building an agent firewall created Agent Evidence Levels (AEL), an open standard that grades how much of an AI agent's recorded actions an outside party can verify without trusting the vendor or operator. The…