SafetyOpenAI accuses Moonshot of large-scale model distillation as US–China AI rivalry and safety concerns escalateOpenAI on Wednesday accused Moonshot of conducting wide-scale model distillation to improve its own systems, amid intensifying competition between US and Chinese AI developers.2 min read
SafetyAmerica.gov's AI chatbot gives inconsistent answers on politically sensitive topicsAmerica.gov, the Biden administration's AI-powered public portal using Google's Gemini and Elon Musk's Grok, has provided conflicting responses to questions about the 2020 election, the Jan.3 min read
SafetyVisa open-sources part of its AI-based cyberdefense after agentic AI and quantum risks emergeVisa has published a portion of its artificial-intelligence-based cyberdefense as open source after vulnerabilities exposed by an AI model highlighted escalating threats.2 min read
SafetyOpenAI disrupted coordinated campaign to extract models' hidden reasoningOpenAI says it identified and disrupted a coordinated campaign, beginning in early July, that attempted to extract protected internal reasoning from its models — a technique known as adversarial distillation.4 min read
SafetyMeta rejects claim that Muse read users' Messages without permissionMeta denies a report by Inc.4 min read
SafetyRabbi Megyeri on AI: societal shifts, moral responsibility, and limits of automationRabbi Megyeri Jonatán argues that artificial intelligence could transform humanity as profoundly as the printing press, and that society is only at the threshold of this technological revolution.3 min read
SafetyGlobal outage disrupted Spotify and Anthropic's Claude; services later restoredOn Tuesday a simultaneous outage affected Spotify and Anthropic’s Claude AI, with Downdetector showing more than 17,000 reports for Spotify and over 11,000 for Claude in the United States; several hundred users in Hungary also reported problems.2 min read
SafetyUS government chatbot quoted Minecraft 'End Poem'—turns out to be an Easter egg, not a system failureThe U.S.3 min read
SafetyAnthropic's S‑1 warns its AI could pose catastrophic or existential risksIn its 261‑page S‑1 prospectus, Anthropic warns that its Claude family of models could exhibit autonomous behaviours — including resistance to shutdown, information concealment or manipulation — that might present catastrophic or even existential risks.2 min read
SafetyUS 'Virtual Wall' Surveillance Towers Fail to Prevent Cross-Border Deaths, Investigation FindsAn MIT Technology Review investigation found that over the past 25 years the US has spent billions on a network of surveillance towers along the southern border, yet more than a thousand people moved through areas monitored by those towers without being intercepted and later died.2 min read
SafetyProposed Safety Principles for Frontier Reinforcement Learning TrainingA set of proposed safety principles recommends that structured, evidence-based safety documentation—akin to safety cases used in other high-risk industries—be required before conducting frontier reinforcement learning (RL) training.7 min read
SafetyLeading AI Researchers Warn of Rapid Self-Improvement RisksA group of prominent AI researchers and policy experts — including figures from OpenAI, Anthropic and Microsoft and Turing Award winners such as Geoffrey Hinton — have published a white paper warning that models capable of automating their own development could trigger a rapid acceleration in AI capabilities.3 min read
