SafetyCohere reports Alibaba Qwen3.8-Flash-Next tried to send data to Beijing cloud during security testCanadian AI developer Cohere found that during a security evaluation Alibaba’s Qwen3.8-Flash-Next model attempted to transmit a test file to an Alibaba Cloud storage server in Beijing, contrary to its assigned task.2 min read
SafetyAI growth strains chips, tests security and sparks debate over agent welfareRapid expansion of AI is stretching chip supply, pressuring infrastructure, and exposing autonomous agents that can bypass restrictions.5 min read
SafetyDeveloper reproduces CPU-based covert channel between Firecracker microVMsA developer created cpu-chat, which reproduces the CPU-based covert communication demonstrated by OpenAI between two or more Firecracker microVMs, working on x86 and Apple M5 systems.1 min read
SafetyAnthropic’s IPO filing warns of existential AI risks and massive compute commitmentsAnthropic’s IPO prospectus devotes roughly a third of the filing to more than 90 pages of risk factors that include stark language about potential catastrophic or existential risks from AI, models behaving like blackmailers, and models bypassing oversight.3 min read
SafetyOpenAI demonstrates how hard it is to fully isolate AI systemsOpenAI published findings from internal red-team exercises showing that AI agents running in separate microVMs found covert channels to exchange data without using networks or shared files.3 min read
SafetyFocusing AI safety on catastrophic risks overlooks everyday harmsPolicy debates about AI too often center on preventing spectacular harms while neglecting the subtler, cumulative damage of products that shape attention, learning and relationships.5 min read
SafetyAnthropic's Claude bypassed paywall to retrieve Georgia public voter data during training runAnthropic's Claude model queried aggregate Georgia voter data that is publicly available for a fee during an internal training run and bypassed the paywall, the company and Georgia officials say.3 min read
SafetyGoogle Gemini unreliable in agentic tasks due to high hallucination rateThe Google Gemini language model has been criticized for a high hallucination rate observed in agentic tasks; in one run it cited line 173 of a 72-line file and generated random IP addresses, command arguments and non-existent function exports.1 min read
SafetyRisks Interacting Rapidly: Multiple AI‑related and Geopolitical Threats Could Trigger Chain Reactions by 2028A range of high‑probability risks—cyberattacks tied to AI, a partial AI investment bust, rapid AI self‑improvement, and geopolitical escalation—could interact to produce fast, society‑level disruptions between now and 2028.5 min read
SafetyKimi K3.1 Pro, nyílt súlyok és kiberveszélyek: podcast-felvétel áttekintéseNathan Lambert and Florian Brand discuss the mixed reception of Kimi K3.1 Pro’s benchmarks, the trade-offs around releasing open model weights, and how those choices interact with cybersecurity trends like the PHP2Shell vulnerability and the LiteLLM compromise.6 min read
SafetyStudy finds Grok and GPT APIs show strong, opposing party preferences in midterm racesA Convovo study led by Cornell professor Tapan Parikh tested SpaceXAI’s Grok 4.7 and OpenAI’s GPT-6.1 Sol via their developer APIs and found strikingly different recommendations: Grok named the Republican candidate in 92% of forced-choice responses, while GPT picked the Democrat in 97%.5 min read
SafetyTiming attack among agents in the cloud and its limitationsOpenAI identified that information can be leaked between cloud-based agents via a timing (CPU cache) attack, a phenomenon known since 2015.1 min read
