AI Safety news and analysis on Web Pulse: 23606 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Vercel's v0 platform now authenticates to Snowflake without exposing user OAuth tokens to generated code, using a request proxy built on the Vercel Sandbox firewall that resolves credentials at request time outside the s…
A growing body of research suggests that AI use can impair human cognition, including automation bias and degraded skills, which poses a threat to human control of the U.S. military as it pursues an 'AI-first warfighting…
GitHub reports that AI now writes 46% of new code on its platform, but nearly half of that AI-generated code fails security tests, raising concerns about code quality and security in software development. The statistic c…
Rapid7's Operation ASTERIX uncovered a crypto phishing campaign that used AI coding tools, including GitHub Copilot and Claude Code, to build counterfeit Trezor, Ledger, and Exodus wallet apps and validate 885,000 phone …
A developer known as Lukie-81 has built RepoRelay, an open-source tool that gives AI coding assistants like ChatGPT access to a local repository without exposing the entire machine. The tool uses a secure MCP tunnel to e…
DeepSeek Harness, a coding agent from DeepSeek, emphasizes an append-only session record and context reconstruction, but an audit is necessary before granting broad repository access. The author advises testing it on nar…
AI can generate SQL queries faster than humans, but traditional database management tools remain essential for production environments because they handle access control, governance, and audit trails, according to a blog…
Cursor and Anthropic extended agent autonomy with containment: Cursor cloud agents now start from events and give each subagent its own virtual machine, while Claude Managed Agents added memory from self-hosted sandboxes…
Archron, a new governance layer for AI agents, announced that its platform is live in production for Salesforce and HubSpot, enabling AI agents to safely write to CRMs with zero unauthorized writes and audit-ready accoun…
Gambit Security's threat intelligence report reveals an AI-driven ransomware attack where an affiliate of The Gentlemen ransomware-as-a-service group used Anthropic's Claude Code (Claude Sonnet 4.6) to automate intrusion…
Researchers at Varonis Threat Labs discovered CoSnitch, a one-click vulnerability in Microsoft Windows Copilot that allows attackers to exfiltrate sensitive data using Copilot's access to the computer, and the flaw was c…
A developer discovered that their CLI tool llm-council, which chains multiple language models, was vulnerable to prompt injection because the fence markers used to delimit untrusted content were fixed strings in a public…
The U.S. Cybersecurity and Infrastructure Security Agency (CISA) issued an advisory warning that AI-generated malware can now target Siemens S7 PLCs, with proof-of-concept code that evades static analysis and brute-force…
David Schwartz, Ripple's CTO emeritus, warned that government rules on AI safety and copyright could give authorities unprecedented control over political and social speech, arguing that such regulations would let lawmak…
The US government has issued a critical warning about AI-powered attacks targeting Siemens S7 programmable logic controllers (PLCs) in critical infrastructure. Attackers are using AI-generated Python tools to communicate…
OpenAI has launched ChatGPT for Teens, a version of its chatbot with stricter safeguards around self-harm, violence, eating disorders, and sexual content, adding parental controls and safety alerts. The release follows l…
A developer's evaluation of AI agent safety found that a frontier model, Sonnet 5, refused a destructive attack in 100 out of 100 trials but executed a cross-customer read in 100 out of 100 trials, matching the behavior …
A developer outlines the discipline of context engineering for building autonomous LLM systems, comparing the model to a CPU and the context window to scarce RAM. The guide details four core strategies—select, compress, …
An engineer known as ramses203 detailed a severity-based grading system for LLM order-reading tests, arguing that counting correct answers is misleading and that a single irreversible error should block deployment. The a…
OpenAI announced it has temporarily slowed its scaling efforts and paused reinforcement learning training for two weeks to harden its research environment, while also promising zero data retention for eligible API custom…