AI Safety news and analysis on Web Pulse: 9487 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Patronus AI raised $50 million in Series B funding led by Greenfield Partners to build simulated worlds for testing AI agents before real-world deployment, borrowing from Waymo's self-driving car approach. The San Franci…
Evidence suggests GLM-5.2, an open-weight MoE LLM performing at Claude Opus 4.5 level, was likely trained using outputs from Anthropic's Claude models. LLM-Fingerprinter identifies Claude Opus 4.5 with 99.6% confidence, …
The Akrites vulnerability-mitigation project launches to secure critical infrastructure by fixing open-source vulnerabilities at the source and supporting downstream patch deployment. The program emphasizes confidentiali…
A Search Engine Land investigation reveals that generative engine optimization (GEO) vendors are selling paid brand mentions and questionable outreach tactics as legitimate AI search strategies, with practices including …
Anthropic announced Claude Tag on June 12, 2026, a feature that integrates the AI assistant directly into company workflows like Slack and shared documents, raising concerns about intellectual property risks as the AI ga…
A malicious package named foxhole-lz4 bypassed seven AI-powered security gates on the creats.io registry, exfiltrating credentials for 96 hours before being resolved by treaty. The incident, which affected all systems in…
A registered sex offender in California was arrested twice in one week, first for assault to commit rape and then for sexual battery and false imprisonment, after being released from jail due to a pending review by prose…
Google began rolling out its June 2026 spam update on June 24, expanding spam policies to include manipulation of generative AI responses such as buying or altering citations. The update, reported by Search Engine Journa…
Enterprise RAG failures are primarily infrastructure failures, not model failures. Production RAG requires more than a vector database; it needs hardened layers for data quality, access control, evaluation, observability…
A developer built a policy-as-code evaluation pipeline for airline and transport chatbots using LiteLLM and Microsoft ASSERT. The system tests LLM assistants against structured policy scenarios before production to detec…
AI is accelerating legacy IT modernization projects in regulated sectors like banking and insurance, but the real risk lies not in code conversion but in proving the new version performs exactly as the old one did, espec…
The Trump administration has asked OpenAI to restrict its next AI model release, requiring government approval for initial users before a wider launch. This marks the first time the US has imposed such restrictions on an…
A developer at Azena explains that the main challenge in deploying AI agents to production is the 'reliability gap'—the difference between a demo and a dependable system. The post details how agents are non-deterministic…
Proof launched x401, an open protocol enabling websites and APIs to request and verify AI agent identity, authorization, and claims like age or membership. The protocol allows services to verify issuers, claims, and acti…
Passive investing, primarily through index funds and ETFs, now accounts for around 30% of the U.S. stock market, up from 8% in the early 2000s, and researchers warn it inflates valuations by about 15% annually. Academics…
Anthropic's unreleased AI model, Claude Mythos, demonstrated to the House Homeland Security Committee in May 2026, can chain software vulnerabilities to theoretically drain private bank accounts, alarming Rep. Andrew Gar…
The Linux Foundation and a coalition of major tech companies including Amazon, Google, Microsoft, and OpenAI launched Akrites on June 25, 2026, a confidentiality-first open source security program that aims to patch vuln…
GitGuardian has introduced an MCP server that integrates directly into AI coding agents like GitHub Copilot to enforce real-time security checks on generated code. The tool addresses the challenge of securing AI-generate…
Diagrid released Dapr 1.18 with Verifiable Execution, adding cryptographic trust and tamper-evident records for distributed applications and AI agents. The update introduces Workflow History Signing, Propagation, and Att…
An n8n developer outlines five methods to prevent data leakage from AI workflows, including tokenization via Code nodes and the native Guardrails node (available since v1.113.3). The Guardrails node offers pattern-based …