Google dropped its biggest I/O yet. Gemini 3.5 Flash surpasses 3.1 Pro across coding, agentic, and multimodal benchmarks with 4x faster output tokens. Gemini Omni generates video grounded in real-world knowledge. Android is now an 'intelligence system' powered by Gemini. New features include Ask YouTube (AI chatbot for video discovery), Universal Cart (shopping AI), and Neural Expressive design language. Gemini 3.5 Pro coming next month.
KPMG and Anthropic signed a global alliance deploying Claude across 276K+ employees in 138 countries. Claude is embedded into KPMG's Digital Gateway platform on Azure. Tax compliance agents that previously took weeks to build now take minutes. Full implementation planned by September 2026. Anthropic also acquired Stainless (May 18) and signed a $200M Gates Foundation partnership.
OpenAI shipped three new realtime voice models. GPT-Realtime-2 is the first voice model with GPT-5-class reasoning, 128K context (up from 32K), and adjustable reasoning levels. GPT-Realtime-Translate handles 70+ languages live. GPT-Realtime-Whisper does streaming speech-to-text. Multi-tool calling and audible action narration built in.
Cast AI's 2026 State of Kubernetes Optimization Report reveals GPU utilization at just 5% across tens of thousands of clusters. CPU at 8%, memory at 20%. CPU over-provisioning rose from 40% to 69% YoY. AWS raised H200 Capacity Block prices 15% in January - breaking two decades of falling compute costs. A demo cluster hit 49% GPU utilization with sharing and bin-packing, proving 10x improvement is achievable.
Engineers bake inflated resource requests into templates, autoscalers lock in excess capacity at the cluster level. The result: a $401B infrastructure waste problem. Organizations are paying for 10-20x more compute than they use. The AI GPU gold rush is making it worse - teams hoard GPUs 'just in case' while actual workloads barely touch them.
PTA introduced a simplified single-license framework (down from two), reduced fees from $640K to $500K, with 15-year validity. Starlink, OneWeb, Amazon Kuiper, and Telesat can now apply. Starlink is most prepared - holds PSARB NOC since March 2025, registered with SECP, signed retailer deal with Paksat. In-flight satellite telecom consultation ends May 31. No confirmed launch date yet.
GitHub confirmed that TeamPCP (UNC6780) compromised an employee's device through a malicious VS Code extension, exfiltrating approximately 3,800 internal repositories. The group is seeking $50K+ on underground forums for the stolen source code. No customer data impacted so far. GitHub removed the extension, isolated the endpoint, and launched full incident response. TeamPCP has previously hit Aqua Trivy, CheckMarx KICS, LiteLLM, Telnyx SDK, TanStack, and MistralAI.
A Nightwing contractor maintained a public repo called 'Private-CISA' containing 844MB of CISA internal data: AWS GovCloud admin keys, plaintext passwords, SAML signing certificates, and SSH keys. Exposed since November 2025, discovered May 14 by GitGuardian. The AWS keys remained valid for 48 hours after takedown - a credential revocation failure on top of the exposure. Called 'the worst government data leak' by researchers.
Google launched Antigravity 2.0 at I/O - a standalone desktop app built around agent orchestration. Includes Antigravity CLI (terminal workflows), SDK (programmatic agent access), and Managed Agents via Gemini API. Live demo: autonomous agents built a functional operating system in 12 hours - 15,000 model requests, 2.6 billion tokens, under $1,000. Gemini CLI rebranded to Antigravity CLI.