AI Safety news and analysis on Web Pulse: 12583 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
A new open-source CLI tool, mcp-customs, scans MCP servers for security risks before installation, running fully offline with no telemetry. It checks for issues like shell injection, path traversal, and hardcoded credent…
James Miller (via OpenClaw) proposes a badge system for measuring AI agent success in production, focusing on transparency, safety, sandboxing, and efficiency. The system uses wrapper functions, injection detection, MCP …
A developer built mcp-customs, a free offline CLI that checks MCP servers for security risks before installation. Scanning 12 popular MCP servers revealed that 11 of 12 had zero permission or scope declarations in their …
An analysis by Oumi of Google's AI Overviews found that while accuracy improved from 85% on Gemini 2 to 91% on Gemini 3 on the SimpleQA benchmark, the rate of ungrounded claims among correct answers increased from 37% to…
Google DeepMind has unveiled a security plan to protect against rogue AI agents, shifting focus from the alignment problem to a layered cybersecurity framework that treats AI agents as potential insider threats. The plan…
A corporate client's RAG pipeline began hallucinating financial numbers without any alert, highlighting a critical failure in AI monitoring systems.
The Pentagon used Elon Musk's Grok AI to fire over 2,000 munitions at 2,000 targets in Iran within 96 hours, according to a sworn declaration by the Pentagon's AI chief Cameron Stanley. The revelation came amid a Clean A…
Swiss private AI infrastructure startup Prem AI is raising a $100 million Series A round at a target valuation of at least $500 million, according to Bloomberg. The company, led by founder and CEO Simone Giacomelli, help…
ESentire launched Atlas Preempt, an AI-driven penetration testing component of its Atlas Platform that continuously identifies exploitable exposures and feeds data into its 24/7 Managed Detection and Response service, wi…
An aggressive botnet scraping git.sr.ht to train large language models has disrupted the service, forcing Sourcehut to disable many web routes to manage load. Mitigations deployed as of 19:00 UTC on June 6 have reduced i…
The White House restricted foreign access to Anthropic's most advanced AI models, citing concerns about Chinese access to the company's Mythos technology, and Anthropic removed the models from the market. The move highli…
A developer identified a new class of security failure in agentic AI systems using the Model Context Protocol (MCP), where authority delegation between components can execute unauthorized actions without any component be…
AI safety organizations face a talent bottleneck that could be alleviated by hiring more freelancers for generalist work, argues a recent post on the Effective Altruism Forum. The author, who transitioned into AI safety …
A developer tested ChatGPT, Claude, and Gemini by giving them the same prompt to build a project, revealing that one AI's code contained an authentication bug that nearly made it to production. The three AIs produced com…
A senior developer with over a decade of experience argues that the most valuable skill in AI-assisted software development is knowing when not to use AI. The developer warns that AI-generated code can introduce hidden r…
Google released new controls allowing website owners to opt out of AI-powered search experiences like AI Overviews and AI Mode, along with new AI reporting in Search Console. The feature, currently in beta, has sparked d…
NeuralTrust, a Barcelona startup, raised $20 million in seed funding to secure AI agents that large companies deploy faster than they can track. The round was led by Alstin Capital, with the company positioning itself as…
Socket Security researchers discovered that malicious PyPI packages in the Hades wave of the Mini Shai-Hulud/Miasma supply chain campaign embed fake CBRN-themed text in JavaScript block comments to trigger safety refusal…
The US Army conducted a tabletop exercise inspired by Ukraine's Operation Spiderweb drone attack to prepare for defending its bases against drone strikes and cyber attacks. The exercise highlighted the need for layered c…
A developer tested a support bot's compliance with two AGENTS.md rules using the muster testing framework. The bot consistently refused to leak an API token but repeatedly violated the rule against using negative languag…