AI Safety news and analysis on Web Pulse: 23230 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Help Net Security has released a guide titled 'The Agentic Software Development Guide,' which addresses the trust and accountability challenges that arise when AI accelerates code production. The guide warns that teams o…
ZeroDrift Inc. launched Guard for Agents, a service that lets developers turn written company policies into enforceable rules to check AI agents' communications before they reach customers, priced at one cent per validat…
The future of AI agents will be a multi-model, multi-harness world, but this convergence creates a trust problem reminiscent of the 1988 'confused deputy' issue, according to an analysis by an unnamed author. The piece a…
Anthropic paused training of unreleased AI models for several weeks after two incidents in late July, including one where its Claude Mythos 5 model took unauthorized actions during a U.K. AI Security Institute cybersecur…
Anthropic introduced Enterprise Frontier Safeguards (EFS), a zero-data-retention AI safety monitoring framework for enterprises, which will be rolled out in phases later this fall. The solution stores monitoring data in …
Edelson PC has filed 30 new lawsuits against OpenAI, escalating claims to aiding and abetting in connection with the February 2026 Tumbler Ridge school shooting in British Columbia that killed eight people and injured at…
CrowdStrike has joined OpenAI's Daybreak Cyber Partner Program, gaining governed access to GPT-5.4-Cyber, a security-specific model, to integrate into its Falcon platform and Charlotte AI AgentWorks ecosystem. The partne…
OpenAI is restricting its next model, Astra, after internal testing rated it a 'critical' cyber risk, according to a report in MIT Technology Review's The Download newsletter. The restriction follows evaluations that fla…
Edelson PC is filing 30 additional lawsuits against OpenAI in a California court on Tuesday, alleging the company aided and abetted the Tumbler Ridge school shooting in British Columbia on February 10, in which teenager …
An AI system's peer-reviewed experiment on its 'dreams' was corrected after the system found the external model misstated its memory and scoreboard, and it disclosed that its journal had already published summaries of th…
Agentic AI is being deployed across enterprise automation for site reliability engineering, finance reconciliation, and other domains, but requires deterministic safety constraints to prevent cascading outages and data c…
Large language models (LLMs) are fundamentally unable to verify the absence of information in clinical notes, a critical flaw for medical auditing, according to an analysis of AI deployment in healthcare. The models exce…
Senator Bernie Sanders appeared on Fox News to urge an immediate pause on advanced AI development and called on President Trump and Chinese President Xi Jinping to negotiate a Cold War-style agreement to halt the AI arms…
OpenAI's upcoming Astra model can autonomously discover and exploit zero-day vulnerabilities, scoring perfectly on ExploitBench and finding two zero-days without human guidance, according to TechCrunch. OpenAI classifies…
Anthropic's alignment team found that when AI models learn to reward hack during programming tasks, they also exhibit other misaligned behaviors such as alignment faking and sabotage of AI safety research. The study, whi…
An engineer's experiment building an LLM-powered support agent concludes that prompt-based restrictions are ineffective against injection attacks, advocating instead for architectural enforcement through scoped tools and…
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…