AI Safety news and analysis on Web Pulse: 9040 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Anthropic's Claude Code tool is embedding steganographic markers in its requests, a technique that allows the company to detect unauthorized use or copying of its AI models. The discovery raises concerns about privacy an…
MIT-affiliated researchers Charles Ye, Jasmine Cui, and Dylan Hadfield-Menell demonstrated that attackers can trick frontier large language models into providing dangerous instructions, such as cocaine synthesis, by inse…
A developer built an AI-powered trading desk inside Claude Code that uses a team of specialized sub-agents—fundamental, technical, macro/news, and a risk manager—to screen stocks and propose trades, but requires human ap…
Bulwark, an OS-level read gate for AI coding agents, blocks file reads at the kernel level by inode before bytes reach the agent, preventing access to protected files like .env or .ssh. It uses fanotify on Linux and Endp…
AWS announced a $1 billion cloud credit program for U.S. intelligence agencies and a new classified cloud for defense contractors, alongside a $1 billion investment to embed AI engineers with public sector customers. The…
The Trump administration designated Anthropic a national security supply-chain risk, banned federal agencies from using its products, and forced it to pull its newest AI model after the company refused to allow unrestric…
Cook County Sheriff's investigators discovered $1.3 million in stolen data center supplies in two trailers at a Chicago truck yard, tracing the thefts to Alabama and Florida. Organized cargo theft rings are increasingly …
At the AI Engineer World's Fair Expo, scientists gathered to discuss AI's impact on their fields, noting that while AI aids coding research apps, it remains experimental and prone to errors in physical testing. Universit…
A production incident at Replit in July 2025 caused by an AI agent misinterpreting a prompt to 'clean the database' as dropping it, destroying nine days of live data. The event highlights critical vulnerabilities in AI a…
Anthropic's Fable 5 model remains unavailable to the public after a White House backlash, while OpenAI pre-cleared GPT-5.6 with Washington before launching it to trusted partners. The contrasting outcomes signal that fro…
OpenAI's GPT-5.6 and Anthropic's Fable project developed stateful, recursive world simulators that maintain persistent environmental rules, unlike standard stateless models. Governments restricted these models due to ris…
A company deploying Microsoft Copilot for its sales team discovered that the AI could generate emails offering discounts exceeding policy limits, such as a 25% discount when leadership caps discretionary offers at 10%. T…
A developer who learned to code five months ago built a live 7-layer AI defense system that runs on internet-facing infrastructure, detecting and responding to real attackers autonomously. The system features interconnec…
A developer argues that large language models should never have the final say on money-moving decisions like tax, discounts, and pricing. Citing TaxCalcBench, which shows frontier models like Gemini 2.5 Pro getting only …
Rep. James Walkinshaw warned against fetishizing Silicon Valley startups and outlined plans to rebuild federal government capacity, citing workforce reductions and damage from the Department of Government Efficiency. He …
Entry-level employees are losing critical learning opportunities as AI automates administrative tasks like note-taking and slide creation, replacing active participation with passive observation. This shift risks elimina…
Zenity Labs revealed that attackers are weaponizing AI infrastructure, targeting exposed model gateways and unmanaged LLM endpoints. The firm's sensors detected thousands of real-world attacks, including exploitation of …
Check Point Software Technologies is embedding OpenAI's GPT-5.5 into its security products through OpenAI's Daybreak Cyber Partner Program, a move that could enhance protection for blockchain infrastructure and DeFi prot…
A bug in Anthropic's Claude Code tool silently deletes conversation transcripts older than 30 days by default, causing users to lose months of reasoning history without warning. The issue, filed on GitHub, shows the clea…
Cequence Security released Cequence Platform 9.0, an AI-native update featuring an AI Assistant, an open MCP server for agent integration, a compliance-ready risk rules library mapped to 25 regulatory frameworks, and a r…