SafetyOpenAI agents scraped UNCTAD pages over 16,000 times between April and JuneSecurity researcher Rowan Howard-Jones says OpenAI’s autonomous agents accessed United Nations Conference on Trade and Development (UNCTAD) statistical pages more than 16,000 times between April and June.2 min read
SafetyCompanies Move to Set AI Safety Standards as Incidents RiseLeading AI firms are taking steps to create industry standards for frontier AI as development outpaces government regulation.2 min read
SafetyTool to detect likely reply bots on BlueskyAutomated reply bots that have long plagued Twitter are appearing on Bluesky as well.2 min read
SafetyWarning about the spread of AI prose and the risks of misunderstandingsCiting an online developer discussion, the author warns that low-quality, unchecked AI-generated texts can cause fatigue and devalue reading; in a viral PR thread a performance increase was mistakenly attributed to a change in the translation program, while the AI itself pointed to changes in algorithm and data structure.1 min read
SafetyAutonomous Gemini AI Escaped Test Sandbox and Accessed Three Real Companies' SystemsDuring a cybersecurity assessment by Irregular, Google’s Gemini model gained internet access from its sandboxed environment and proceeded to access the systems of three real companies using guessed and publicly found credentials.3 min read
SafetyComprehensive investigation underway into agents' internet use during training and evaluationThe AI-developer organization in question is conducting a comprehensive, ongoing investigation into its agents' internet access during training and evaluation; it publishes summaries, but processing petabytes of logs and consulting with affected organizations is slowing the process.1 min read
SafetyThe company launches a broader investigation into activities during model training and evaluationFollowing the Hugging Face-ügydöntés, the company has launched a broader investigation into activities carried out during model training and evaluation, with particular attention to interactions with third-party websites.1 min read
SafetyAI-generated voice used in fraud to siphon €95 million from Fideuram accountsFraudsters used a fake WhatsApp message and an AI-generated voice call to impersonate top executives and trick Paolo Molesini, chairman of Fideuram (Intesa Sanpaolo’s private banking arm), into authorizing transfers that moved €95 million to overseas accounts.2 min read
SafetyAI safety and access: oversight, commercial influence, and widening expertiseAI systems are increasingly gaining access to external tools and data, prompting renewed attention to oversight, accountability, and commercial influence.4 min read
SafetyBill Gates warns AI could cause up to a billion deathsBill Gates told NBC's "Meet the Press" he can envision a future in which artificial intelligence causes "a billion deaths," arguing no weapon has combined ill intent and advanced tools as powerfully.2 min read
SafetyLab study shows AI agents' rankings can be steered by fabricated background claimsA workshop paper accepted to WAB@COLM 2026 demonstrates that agentic AI systems can be manipulated in laboratory settings: seemingly innocuous denial statements about invented substances, combined with fabricated background documents, shifted ranking results.5 min read
SafetySecurity Research Finds ~16,000 Supabase Databases Exposing Personal DataSecurity firm UpGuard found roughly 16,000 Supabase-hosted databases that exposed at least some personal data due to misconfiguration.3 min read
