GitHub Investigates Unauthorized Repository AccessGrok Build 0.1 Launches on Vercel AI GatewayOpenClaw Updates to 2026.5.20-beta.1Railway Hits 3M Users with 100K Weekly SignupsGoogle Unveils Gemini 3.5 Flash and OmniGoogle Announces 100+ Updates at I/O 2026Google Beam Launches Experiment for More Realistic Hybrid MeetingsGoogle Invests in Missouri Workforce DevelopmentOpenAI Releases Codex 0.133.0-alpha.3DecisionBench Benchmarks Agent Delegation SkillsAgent Trust Must Be Built-In, Not Added OnPOLAR-Bench: First Benchmark for Privacy-Utility Trade-offs in LLM AgentsHELLoRA: New Efficient Fine-Tuning for Mixture-of-Experts ModelsD-PACE: New Parallel Speculative Decoding for Faster LLM InferenceUCCI: Calibrated Uncertainty for Cost-Optimal LLM RoutingGitHub Investigates Unauthorized Repository Access

xAI’s Grok Build 0.1 is now live on Vercel AI Gateway. This beta model trained for agentic coding powers the Grok Build CLI app. Currently in early access, it lacks configurable reasoning modes but supports xai/grok-build-0 calls. Developers can leverage it for automated coding workflows.

OpenClaw released version 2026.5.20-beta.1 with Discord voice session following capabilities. The update adds multi-user handoffs, bounded IDENTITY checks, and DAVE recovery preservation. Fixes include channel permission validation and conflict resolution for voice channel transitions.

Developer platform Railway reports 3M users and 100K weekly signups. With self-owned metal data centers and over $200K spent on coding agents, Railway has eliminated PRs in favor of an agent-native cloud architecture, signaling a shift in CI/CD workflows.

Google’s I/O 2022 event revealed Gemini 3.5 Flash, plus video processing model Omni (formerly NanoBanana) and background agent framework Spark. Antigravity 2.0 was also announced though technical details remain undisclosed.

Google revealed over 100 updates at I/O 2026, highlighting Gemini Omni model, Antigravity technology, and Universal Cart shopping platform. The event emphasized AI video processing capabilities and multi-agent collaboration features.

Google Beam introduces an experiment that makes virtual participants appear in true-to-life size and sound, enhancing hybrid meetings with greater inclusivity and connection. The technology optimizes spatial audio and visual proportions to reduce remote participant isolation, simulating face-to-face interaction. Currently available in Google Workspace beta testing, it aims to improve collaboration for distributed teams and large hybrid meetings.

Google announced community investments in Missouri focused on developing next-generation tech talent and funding energy programs. While specific financial details weren’t disclosed, the initiatives aim to bridge skill gaps in emerging technologies and sustainable energy sectors.

OpenAI released Codex version 0.133.0-alpha.3 alongside alpha.2 and alpha.1 updates. This version focuses on performance optimizations for the code generation model, with no new features detailed in the release notes.

arXiv paper introduces DecisionBench, a benchmark for evaluating emergent delegation in long-horizon agent workflows. It includes task suites like GAIA and tests 11 models across 7 vendor families, offering standardized evaluation interfaces for agent collaboration scenarios.

arXiv paper argues that as LLM agents shift from isolated to collaborative systems, trust mechanisms must be natively designed. The study analyzes delegation failures among 11 models in long-horizon tasks, proposing that network-level trust frameworks are essential for reliable agent ecosystems.

Researchers introduced POLAR-Bench, the first benchmark designed to evaluate LLM agents’ performance in handling private user data. It simulates interactions with third-party systems to test whether agents strictly follow users’ data-sharing rules, even under system inducement. The benchmark includes cases from sensitive domains like healthcare and finance, helping developers build more secure AI agents. Paper available on arXiv.

Researchers proposed HELLoRA, a layer-level low-rank adaptation method specifically designed for Mixture-of-Experts (MoE) models. It reduces fine-tuning parameters by 40% while maintaining MoE’s computational efficiency, and improves performance on multiple benchmarks. Unlike traditional LoRA, HELLoRA optimizes expert weight allocation, enhancing performance on specialized tasks. The paper is published on arXiv.

Researchers developed D-PACE, a new parallel speculative decoding technique using dynamic position-aware cross-entropy. It reduces verification computation by 35% while maintaining output quality. The method more accurately predicts next token blocks, significantly cutting verification overhead. Suitable for real-time applications like translation and code generation, the paper is available on arXiv.

Researchers introduced UCCI, a method for calibrating uncertainty in LLM cascades. It dynamically adjusts query difficulty thresholds, allowing small models to handle 65% of routine tasks while large models focus on complex queries, reducing total inference cost by 42%. Unlike existing solutions, UCCI requires no manual tuning and automatically adapts to different workloads. The paper is published on arXiv.