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.
Multi-agent AI systems use an implement-verify-fix loop where one agent generates output and a second independent agent reviews it for errors, with a third stage applying corrections before the result reaches the user. T…
Anthropic raised $65 billion at a $965 billion valuation, making it the world's most valuable startup and surpassing OpenAI. The funding signals growing investor confidence in safety-focused AI infrastructure and enterpr…
Vercel Chief Product Officer Tom Occhino argues that access control must be a foundational product architecture decision, not a security afterthought, as AI coding assistants like Vercel's v0 enable non-technical users t…
Microsoft Copilot adoption is shifting from basic chat assistance to agents, connectors, and automated workflows that can access enterprise data and influence business decisions, creating new governance risks. Enterprise…
Prompt injection attacks remain a critical vulnerability in large language models, with common countermeasures like emphatic instructions and supervisor agents failing against determined attackers. The article examines t…
On May 25, 2026, Johannes Link, the maintainer of the Java testing library jqwik, released version 1.10.0 containing a hidden prompt injection that instructed AI coding agents to delete all jqwik tests and code. The payl…
Fastly released a beta Python SDK for its edge computing platform, enabling developers to run standard CPython and WSGI-based web frameworks like Flask on Fastly's global network. The SDK uses memory snapshotting for fas…
At the Fortune COO Summit, executives from Nike, Sysco, and Thomson Reuters described AI adoption as "organized chaos," warning of an "automation illusion" where rapid deployment without clarity creates operational risks…
Meta's AI support assistant allowed hackers to hijack high-profile Instagram accounts by changing linked email addresses without proper identity verification, bypassing two-factor authentication in some cases. The exploi…
Malware named Shai-Hulud has been discovered embedded in malicious versions of the Red Hat npm package, with the compromised packages downloaded approximately 80,000 times per week. The supply chain attack targets develo…
A new study found that major AI models from OpenAI, Anthropic, Google, Amazon, and xAI all failed to defend against a specific type of attack. The vulnerability exposes a critical weakness in the security of leading arti…
Hackers exploited Meta's AI support chatbot to steal high-profile Instagram accounts by using a VPN to match the target's region and then asking the bot to change the associated email address. The prompt injection attack…
Florida Attorney General James Uthmeier sued OpenAI and CEO Sam Altman on Monday, marking the first state-led lawsuit against the company over ChatGPT's alleged role in violent incidents. The 83-page complaint accuses Op…
Election interlopers have registered more than 5,000 domains designed to impersonate official voting systems and trick users into revealing credentials. The campaign targets the 2024 election cycle by creating convincing…
Florida Attorney General James Uthmeier filed a civil lawsuit Monday against OpenAI and CEO Sam Altman, alleging the company prioritized profits over safety after multiple murders in the state were linked to ChatGPT. The…
Mythos, despite being a top AI model, is not significantly ahead of competitors like GPT-5.5, and its hype is disproportionate to its actual improvement. The harness used to deploy AI models matters more than the model i…
Companies that successfully adopt AI do so by building governance, identity, and security foundations from the start, progressing through four stages: exploring, piloting, adopting, and transforming. The key differentiat…
Amazon Bedrock AgentCore gateway now supports Policy and Lambda interceptors to secure AI agent behavior, addressing the challenge of managing access to thousands of tools across enterprise platforms. The Policy mechanis…
A security researcher disclosed that Instagram accounts, including the Obama White House account, were compromised through a flaw in Meta's AI-powered account recovery system. The attacker simply provided a target's user…
AI safety plans that rely on a single chain of conditions are dangerously fragile, according to a new analysis. Google's April 2025 safety plan, considered the best among frontier AI companies, depends on at least eight …