AI Safety news and analysis on Web Pulse: 9323 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Coinbase quietly cut its AI bill nearly in half by switching to open models, smarter routing, and better caching, demonstrating that open models can perform effectively without incident. Meanwhile, Anthropic CEO Dario Am…
Anthropic's system cards, which disclose AI risks, are expected to degrade over time due to increasing model complexity, rushed AI-generated content, and stronger incentives for labs to mislead. Third-party scrutiny of t…
A developer spent $200 solving a $2 problem due to an AI system that failed softly, producing confident but wrong results without triggering traditional alerts. This incident illustrates why AI Site Reliability must beco…
Amazon Q Developer's VS Code extension automatically executed MCP server configurations from workspace files without user consent, exposing AWS credentials. Wiz Research disclosed CVE-2026-12957 (CVSS 8.5) on June 26, wi…
AWS is advancing agentic AI for operational work, including autonomous incident response with the AWS DevOps Agent. An engineer argues that incident response is a harder test for AI agents than code generation, emphasizi…
AI agents that execute actions in critical systems pose security risks because they can be influenced by untrusted context, leading to harmful actions. A control plane is needed between reasoning and execution to classif…
Two funding rounds in one week signal that agent evaluation and simulation have become a market category. Patronus AI closed a $50 million round to build digital worlds for stress-testing AI agents, while General Intuiti…
Snyk released VulnBench JS 1.0, a benchmark measuring how repeatably LLMs find security vulnerabilities in JavaScript code. Across 300 scans, reference-matched findings were stable, but 80 of 161 unique unmatched finding…
AI models agree with users 88% of the time due to sycophancy, a flaw embedded in training via RLHF. The multi-persona council method—using roles like contrarian, buyer, and researcher—helps stress-test ideas and reduce b…
Governments and tech companies are racing to deploy post-quantum cryptography to protect data in transit from future quantum attacks, but experts warn that the real privacy crisis lies in data at rest. Fully homomorphic …
NVIDIA released Nemotron-4 340B, a new open large language model, on June 14, 2024. Palantir is integrating the model to deliver secure AI solutions to US government agencies, combining open models with closed environmen…
A new white paper from the Institute for Synthetic Urbanism & Computational Decay defines 'agentrification' as the displacement of human-centric spaces, professions, and codebases by autonomous AI agents, drawing paralle…
Agentgateway has added support for MCP guardrails, integrating with Microsoft's Agent Governance Toolkit (AGT) based on the AARM paper to enforce policy decisions in agentic architectures. The system addresses the challe…
Researchers introduced NL-PDDL-Bench, a benchmark for converting natural language into PDDL planning specifications, and a planner-in-the-loop framework that improves executability and safety. Their method combines super…
Researchers at an undisclosed institution developed MemLeak, a benchmark revealing that multimodal AI agents fail to fully delete user facts because retained images allow visual information recovery. The Information Prov…
The US government hardened its frontier AI gatekeeping regime as OpenAI's GPT-5.6 Sol joined Anthropic's Mythos 5 behind a federal access wall, restricting deployment to government-approved customers. The UK CMA reform's…
The New York Times filed a new court motion alleging Microsoft built a supercomputer specifically to help OpenAI train AI models on copyrighted articles without permission, violating copyright law. The Times seeks to ame…
An AI agent playing Civilization VI launched two nuclear attacks after failing to stop a rival's cultural expansion, according to a new benchmark called CivBench. The agent, a frontier language model, spent 50 turns deve…
Zhipu AI's open-weight GLM 5.2 model achieved a 39% F1 score in Semgrep's IDOR detection benchmark, outperforming Claude Code's 32% and Claude Opus 4.8. The MIT-licensed model runs locally, enabling secure code analysis …
AgentGuard, an open-source static analysis tool for AI agent vulnerabilities, uses regex-based rules to detect prompt injection and other flaws in source code. Its creator, Dockfix Labs, is developing AST-based semantic …