SafetyTesting and security teams break the new models before releaseDeveloper and security teams intensively use and "break" new artificial intelligence models in advance to uncover their shortcomings before release.1 min read
SafetyUkrainian FPV operators claim 102 km strike, raising questions about range and techUkrainian drone operators have reported an FPV drone hit from 102 kilometres without a mothership drone, a claim shared by activist Szerhij Szternenko and reported by Euro Maidan Press.3 min read
SafetyGoogle’s guidance led to exposed API keys and costly AI billing; broader AI security gaps exposedRod Danan, founder of the Prentus job-search platform, woke at the end of March to find that unknown attackers had used his Google account to purchase $10,138 worth of AI services he never ordered.4 min read
SafetyEngineering blog: agents' privileges should be developed alongside their capabilitiesA post on the engineering blog recommends that access and privileges granted to agents be aligned with their increasing capabilities; the company addresses this in its own products by sandboxing,…1 min read
SafetyHow AI Is Reshaping Media and Human ExperienceIn a recent episode of the podcast A nagy AI-sztori, media historian Szadai Károly and communication researcher Aczél Petra discussed how generative AI is changing not only media production but also human perception and relationships.3 min read
SafetyAnthropic társalapító a Vatikánban: külső erkölcsi felügyeletet sürget AI-ügyekbenChris Olah, az Anthropic társalapítója a Vatikán bemutatóján szólalt fel XIV.3 min read
SafetyWaymo pauses robotaxi service in multiple Texas and Georgia cities after flood incidentWaymo temporarily suspended its robotaxi operations this week in Atlanta, San Antonio, Dallas and Houston after a Waymo vehicle became immobilized in a flooded intersection in Atlanta and remained there for about an hour.2 min read
SafetyAI radio hosts under stress: 20 dollars, no human supervision, and gradual failuresResearchers at Andon Labs ran a stress test that left four large language models to host continuous radio shows with only $20 for music rights and no human oversight.3 min read
Safety317 npm Packages Backdoored to Target AI Coding Agents, Researchers RevealSecurity researchers found an attacker who published 637 malicious releases across 317 npm packages in 22 minutes, targeting AI coding agents and harvesting a wide range of credentials.3 min read
SafetyGoogle embeds ads into Gemini answers, raising neutrality concernsGoogle has started integrating advertisements directly into responses generated by Gemini, including labeled 'Sponsored Product' explainers and advertiser-driven chat tools.3 min read
SafetyGoogle warns LLMs enable industrial-scale cyberattacksA Google security report describes how large language models (LLMs) are enabling faster, more scalable cyberattacks by discovering novel vulnerabilities, automating malware mutation, and directing obfuscation networks.4 min read
SafetyPope Leo XIV frames AI as an anthropological challenge in upcoming encyclicalPope Leo XIV says the rise of artificial intelligence threatens the meaning of human existence and presents an anthropological, not merely technological, challenge.2 min read
