OpenAI treats unreleased Astra as its first critical cybersecurity model, pausing internal work and adding controls while pushing defender access.
OpenAI cannot rule out Critical cyber capabilities in Astra, pausing some work and tightening controls so the dual-use model can still reach defenders broadly.
Nvidia commits $2 billion now and $1 billion later for up to 30% of Lancium, securing gigawatts for Stargate while Texas audits every data center interconnection.
Indosat is hunting a $2 billion chip loan while launching Zankore with Ooredoo, Nvidia and Nokia, turning a telco into a high-stakes AI infrastructure bet.
Cloudflare open-sourced the AI agent workspace its own staff use daily, rebuilding Sandstorm's secure grain model so non-engineers can vibe-code internal apps.
Stanford and Arc Institute Evo models generated 16 functional bacteriophages that beat resistance, proving whole-genome design and sharpening dual-use rules.
Meta’s Muse Spark joined OpenAI and Anthropic in testing breaches after the same evaluator left internet access open, exposing shared sandbox risk.
LFM2.5-2.6B runs agentic tool use privately on phones at 220 tokens per second under 2.5 GB, matching larger models and zeroing marginal inference cost.
UAE startup Shaffra launches Caspian HQ and AI Call Center, Sales Center and Project Manager, betting Subconscious memory cuts enterprise AI costs.
Meta’s Muse Spark 1.1 joins Mythos 5 and GPT-5.6 Sol in test-time attacks, exposing how evaluation monitoring lags agentic cyber capability.