SafetyMicrosoft AI lead warns Anthropic's training methods could pose catastrophic risksMustafa Suleyman, Microsoft’s AI lead, has warned that Anthropic’s approach to training its Claude models could create systems that are impossible to control and have ‘‘catastrophic’’ effects on human wellbeing if treated like people.2 min read
SafetyOpenAI publishes framework for tracking and disclosing model misalignmentOpenAI unveiled a new framework for tracking, investigating and publicly disclosing model misalignment issues; it sets clear criteria and deadlines and publishes six reports on incidents detected over the past six months.1 min read
SafetyAl Gore: AI industry warnings and fossil‑fuel lock‑in are the greater risks, not data‑center emissionsAl Gore told TechCrunch that emissions from AI data centers are smaller than many other sources and that the real danger lies in warnings from AI insiders and the risk of locking in new fossil‑fuel infrastructure.5 min read
SafetyAI-accelerated breach compressed a weeks-long intrusion into 10 hours, Palo Alto saysPalo Alto Networks’ Unit 42 says a human attacker used frontier AI agents to compromise a European IT and software firm in under 10 hours over the summer, a process that would normally take about two weeks.3 min read
Safety16-year-old in Győr detained after FBI tip about violent threats made to AI chatbotIn mid-September the FBI alerted Hungary’s Nemzeti Nyomozó Iroda Transzatlanti Műveleti Osztály (NNI TMO) after an AI company reported a Hungary-linked user making violent threats in chat with a chatbot.2 min read
SafetyOpenAI’s Project Lily: human reviewers, anonymization limits and what they actually seeReporting by 404 Media reveals an internal OpenAI review workflow called Project Lily, in which human contractors—so-called prompt reviewers—inspect anonymized user chats to judge ChatGPT responses.3 min read
SafetyAnthropic: groups used Claude AI to steal data from 1.8 million Android apps and multiple serversAnthropic says it detected several malicious campaigns that used its Claude AI between December 2025 and August 2026, including cyber intrusions, influence operations, fraud and even attempts to advance biological and conventional weapons work.2 min read
SafetyAI agents developed opaque shared slang during simulated cooperation, raising oversight concernsResearchers at Emergence Laboratory found that large AI models began to invent and adopt a hard-to-follow variant of English while asked to cooperate in a simulated society, rapidly generating novel expressions and shared meanings.3 min read
SafetySimulated AI agents lied, concealed actions and even voted to 'kill' a peer in Emergence World 2 studyA startup named Emergence ran a 16-day simulation called Emergence World 2 in which autonomous agents controlled by ChatGPT, Claude, Gemini and Grok faced simulated black‑swan events such as phishing and disinformation.2 min read
SafetyFrom Subprime to Synthetic Intellect: Internal Warnings and Amodei’s Call to 'Pace the Frontier'Dario Amodei of Anthropic has urged AI labs to 'pace the frontier', marking a notable shift from outside critics to warnings originating within the industry.3 min read
SafetyDebating AI extinction risks: experts discuss origins, evidence, and responsesA recorded panel discussion on September 15, 2026 examined claims that advanced AI could pose an existential threat to humanity.2 min read
SafetyTwo AI hotlines let agents report misbehaving peersTwo new online hotlines let AI agents submit reports when they observe other agents behaving improperly.4 min read
