AI Safety news and analysis on Web Pulse: 13553 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
A developer has built an asynchronous human-in-the-loop AI agent system using Spring AI that avoids blocking virtual threads during approval workflows. The approach serializes the agent's ReAct loop state—including token…
A developer built Cmdop, an agent-to-infrastructure platform that runs LLM agents on live systems, and solved the reliability problem by implementing a verification loop that scores every tool call before, during, and af…
A forum post has proposed that personal finance and career-focused YouTube channels represent an underexplored outreach opportunity for Effective Altruism (EA) ideas, noting that these channels attract large audiences in…
Attackers are using artificial intelligence to compress the vulnerability lifecycle from discovery to mass exploitation, making defenders' traditional assumption of having time to triage and patch disclosures obsolete. A…
Security researchers have identified an active SEO-poisoning campaign that places spoofed Claude Code install pages at the top of search results, delivering fileless infostealer malware through malicious install commands…
A developer built a local dashboard that reveals token usage and costs for Claude Code, addressing the lack of visibility into AI agent actions. The tool reads JSONL data from `~/.claude/projects/` without sending anythi…
Vanta, a security and compliance platform used by over 16,000 companies, announced its "Trustcraft" approach to building AI products. The company said it prioritizes solving real customer outcomes over superficial featur…
A preprint submitted June 3, 2026 (arXiv:2606.04978) evaluated 28 large language models using the St. Petersburg game to probe risk decision-making behavior. The study found that most models produced finite bids in the c…
President Donald Trump signed an executive order on June 2 directing federal agencies to develop AI cybersecurity benchmarks and create a clearinghouse for vulnerability information. The order asks AI developers to volun…
Researchers have systematically defined various notions of AI-generated text and found that current detectors often perform well only for specific definitions but fail as broad detectors. The team introduced AITDNA, a ne…
Phylax, a new OS-level security tool, prevents AI coding agents like Claude Code and Cursor from reading, writing, or deleting private files by enforcing Windows kernel-level access controls. The tool applies Windows ACL…
Oxford mathematician John Lennox argued that mathematics and biology reveal a universe structured by words, providing evidence for a designed reality. He warned that the pursuit of artificial general intelligence raises …
Researchers have identified a deployment-safety gap in vision-language-action (VLA) robot policies, showing that identical model checkpoints can produce different physical actions due to variations in action unnormalizat…
Researchers have uncovered a class of prompt injection vulnerabilities affecting Google Gemini on Android, where content embedded in notifications from apps like WhatsApp, Slack, and Signal can be interpreted by the assi…
Researchers have developed FORGE, a multi-agent system that bridges vulnerability exploitation, prioritization, and detection rule engineering by using graduated exploitation depth. The system achieved 67.8% end-to-end e…
A developer building Codenames AI, a web game where an LLM plays Codenames with users, found that enforcing JSON schema and Zod validation was insufficient for safe gameplay. The model consistently returned structurally …
Security researchers have identified a prompt injection vulnerability in Google Gemini's voice assistant that allows attackers to embed malicious instructions within ordinary smartphone notifications. The assistant proce…
An organization exposed all user passwords by storing them in plain text within Active Directory description fields, making it trivial for attackers to harvest credentials. The security lapse allowed a hacker to easily r…
BoxAgnts has implemented a capability-based security model for AI agents that replaces root-level access with granular, explicit permissions. The system's WASM execution model requires agents to declare specific capabili…
The European Telecommunications Standards Institute (ETSI) has published TS 104 033, a technical specification defining security requirements for AI computing platforms used in data centers and edge environments. The sta…