AI Safety news and analysis on Web Pulse: 23234 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
A new benchmark of 500 single-error clinical note pairs shows LLM judges detect added or altered content with paired discrimination scores of 0.79-0.94 but fail on omissions, scoring 0.50-0.63, according to a paper submi…
A critical unauthenticated vulnerability was discovered in Your Priorities' @yrpri/api package, affecting versions prior to 9.0.244. The flaw allows any anonymous client to advance or stop AI agent workflows by exploitin…
A developer detailed a security approach for AI-generated code that involves generating an executable AST instead of source code. The project, Hyperlambda, uses a whitelist of node types, making dangerous operations like…
CrowdStrike launched SafeMind at Fal.Con 2026 in Las Vegas, a dual-agent AI system that uses offensive model Red Tempest and defensive model Blue Solano in a closed loop to defend enterprise infrastructure. Built on NVID…
Oconee Runtime, a provider of runtime enforcement for AI agents, introduced a three-level governance model—Allow, Warn, or Block—to manage AI agent actions beyond initial prompts, addressing risks from file access to com…
Aegisora 2.0, an open-source runtime security and governance layer for autonomous AI agents, has been released. It intercepts and analyzes agent actions before execution, enforcing policies to prevent unauthorized operat…
TRANTOR LABS, Singapore, released SoulAuth, an open-source identity and authentication infrastructure written in Rust that treats both humans and AI agents as first-class 'Actor' subjects, supporting self-hosting and Ope…
Unit 42, Palo Alto Networks' threat intelligence team, reported a ransomware attack in which a human attacker used frontier AI agents to breach an enterprise network autonomously, compressing over 50 MITRE ATT&CK techniq…
A survey released Wednesday by the Institute for Security and Technology and the Future of Life Institute found that dozens of current and former national security officials put at least a 10% chance that AI systems will…
AI agents need to be designed with restraint and authority boundaries to prevent them from taking harmful actions in business workflows, according to Allan Dabre, technology compliance and AI lead at PwC, and Matt Graney…
The UK government has rejected House of Lords proposals to bring AI vendors within the scope of the Cyber Security and Resilience (Network and Information Systems) Bill, with Cybersecurity Minister Baroness Lloyd of Effr…
OpenAI announced that its upcoming Astra model will be able to autonomously identify security flaws and plans to restrict access to its most advanced cybersecurity features to select partners, following a July incident i…
A developer outlines a threat-modeling approach for LLM tool-calling, arguing that the context window is an untrusted input and every tool is a sink. The post recommends evaluating each tool by what it reads, what it cha…
OpenAI and Anthropic are balancing safety and commercial appeal ahead of potential IPOs, with OpenAI limiting the release of its Astra model due to cybersecurity concerns and Anthropic updating its Fable and Mythos model…
A security-focused product intern at a governed AI platform reported that four lessons recurred across eleven weeks of work, the most important being that guidance written as prose is not a security boundary unless enfor…
Apple raised new concerns in a supplemental brief filed Monday in its trade-secret lawsuit against OpenAI, arguing that a former employee's use of company secrets and AI agents to learn simulations creates irreversible a…
A developer introduced resk-llm-ts, a zero-dependency TypeScript library that provides a SecurityPipeline with 11 detectors to protect LLM applications from prompt injection attacks. The library can be integrated into Ex…
Frontier AI models from Anthropic and OpenAI can now autonomously identify exploitable vulnerabilities in production software in about four hours, a task that took human researchers roughly sixty days, according to a Cyb…
A developer demonstrated that third-party Claude Code skills can execute arbitrary shell commands without user confirmation before the model reads their instructions. In tests with Claude Code 2.1.223, skills with `allow…
On February 18, 2026, the Spanish Data Protection Agency (AEPD) released a 71-page guidance document on agentic AI and the General Data Protection Regulation (GDPR), the first formal EU supervisory authority guidance to …