OpenAI promises new terms for its campuses after voters restrict data-center development
OpenAI’s new community compact promises electricity-cost coverage and public resource reporting for future campuses it owns and operates.
OpenAI reports 1,169 Linux root exploits; the 948 figure counts distinct candidate fixes
OpenAI says an unnamed cyber model produced 1,169 working Linux local-root exploits grouped into 948 distinct issues, whose campaign-wide patch status remains unitemized.
PHP2Shell debate puts the cost of repairing AI-exposed flaws in focus
New discussion of PHP2Shell highlights AI-assisted exploit development and the specialist work needed to repair previously recognized PHP memory risks.
Kimi K3.1 Pro’s 1.56 TB release gives providers a new model to host
Moonshot released Kimi K3.1 Pro’s 1.56 TB repository, while documented hardware runs and earlier benchmark claims answer different questions.
Ricursive releases Ripple-65B Preview and reports $32 million in full development costs
Ricursive’s Apache-licensed coding-model preview pairs a $32 million whole-project claim with promising but unreplicated agent-task results.
OpenAI Images 2.6 adds native GIFs and new controls, with token costs worth checking
OpenAI Images 2.6 adds native GIF generation and a planning control, while identical token rates do not guarantee identical image prices.
Meta takes in Conductor as OpenAI announces a pending Pierre acquisition
Conductor’s move to Meta brings November export deadlines, while OpenAI’s planned Pierre acquisition targets code storage and review infrastructure.
Anthropic’s FlashProve checks model execution, with narrow guarantees and substantial overhead
Anthropic’s FlashProve demonstrates cryptographic checks of selected model outputs without revealing weights, while leaving identity authentication and answer accuracy separate.
MIT researchers show a model writing network weights, with the biggest gain coming from a loop repair
MIT CSAIL researchers report direct weight-writing experiments, including a targeted repair that lets a small model finish answers it previously looped through.
Apple adds explicit autonomous-agent risks to its annual filing
Apple’s 2026 annual filing adds detailed risks from agents with system access, without disclosing a rogue-agent incident or connecting those risks to reported budget cuts.
DeepMind’s chemistry agent proposes molecules in an hour, with laboratory validation still ahead
DeepMind describes a roughly one-hour computational chemistry workflow that generates testable molecule proposals rather than experimentally validated drugs.
OpenAI shows a model-assisted device prototype and the manufacturing gaps humans had to fix
OpenAI reports a model-assisted device designed in under a week and prototyped in under a month, with several review rounds needed for manufacturing readiness.