Anthropic launches Claude Sonnet 5.5, claiming 30%+ faster performance and up to 30% lower cost than Sonnet 5.
World Labs announces it is joining AMD, a major acquisition linking spatial-intelligence research to AMD’s AI hardware push.
NVIDIA launches Open Agent Safety Platform, combining OpenShell, BlueField monitoring, and hardware isolation for agents.
Reports an AI agent finding and exploiting a KVM use-after-free to escape a heavily hardened Google hypervisor.
Flags a Pachocki, Hinton, Bengio, and Clark paper on intelligence-explosion risks from automating AI R&D.
Artificial Analysis launches its Cyber Index and alliance to benchmark AI agents on enterprise vulnerability discovery and patching.
Perplexity shares 108 sandbox-escape red-team runs across nine models, plus patched network-access bypasses.
Meta researcher introduces RL-XAR, a rubric-learning approach claiming major reductions in low-quality AI-generated writing.
OpenAI engineer previews a Pro usage-accounting overhaul alongside 50% price cuts for GPT-6 Sol and Luna.
Highlights that OpenAI API access may allocate substantially more reasoning budget to the same models than Codex does.