本周 AI 回顾 — 2026年7月27日 – 8月2日

Google DeepMind announced that its latest AI model, Gemini Robotics 2, can now control the entire body of humanoid robots. This new version supports "whole-body motions" from feet to fingertips, a significant advancement from the previous model which only focused on controlling the upper body. This development marks a step forward in robotic control capabilities.

Google announced that its internal AI tools helped patch more security flaws in the Chrome browser in June than in the past two years combined. A chart published by Google, as part of a white paper on using AI to find and fix flaws faster, illustrates this exponential increase. Chrome’s 126 was released in June 2024, with Chrome 149 and 150 released last month, each version being a "milestone."

The Gemini Robotics team developed "Gemini Robotics 2," a system designed to bring whole-body intelligence to robots. This initiative involved a large team of researchers and engineers, including Abhijit Ogale, Abhishek Jindal, Adil Dostmohamed, and many others from DeepMind. The project aims to advance robotic capabilities by integrating sophisticated intelligence across the robot's entire physical structure.

OpenAI has open-sourced @openai/codex-security, a CLI and TypeScript SDK designed to identify, validate, and remediate security vulnerabilities in code. This tool allows users to scan repositories, review changes, track findings, and integrate security checks into their CI pipelines. It requires Node.js 22+, Python 3.10+, and access to Codex Security, with detailed documentation available for setup and usage.

Kimi-K3 on HuggingFace presents various benchmark results, including Coding and Agentic capabilities. For Coding, benchmarks like DeepSWE, ProgramBench, Terminal-Bench 2.1, FrontierSWE, SWE-Marathon, PostTrainBench, MLS-Bench-Lite, SciCode, and Kimi Code Bench 2.0 are listed with their respective scores. Agentic capabilities are evaluated using BrowseComp. Additionally, PerceptionBench is mentioned as an in-house benchmark focusing on atomic visual perception capabilities.

A local, zero-cost merge queue for parallel Claude Code agents is introduced, designed to serialize agent operations to prevent push races, redundant heavy builds, and shared-resource test flakiness. The system supports configuration options such as `branchPrefix`, `worktreeSuffix`, `portBase`, `integrationBranch`, and `productionBranch`. It also allows defining `protectedBranches`, `regenerableFiles`, `symlinks`, `buildOutputDirs`, and a `checkCommand` for gating landings, with `checksRequired` to enable or disable these checks. The project is open-source under the MIT license.

ORCA-bench is a new benchmark designed to evaluate language model agents in a production-fidelity oncall setting for root cause analysis (RCA). It uses a live OpenTelemetry-instrumented microservice system with six days of metrics, logs, and traces, and 1,079 RCA tasks. Expert SREs curate ground-truth symptoms. The best agents achieved only 25.3% RCA Accuracy on Medium-difficulty tasks and 10.0% on Hard tasks, even with Claude Fable 5. This indicates a significant gap before these agents can be safely entrusted with production reliability.

A 2025 paper from Northeastern University and the University of California, Berkeley found that 30% to 60% of the "thinking steps" in frontier open-source LRMs had "minimal causal impact" on their answers to math questions. Removing these steps barely affected performance, suggesting that chain-of-thought prompts may not always be linked to the final output. While some state-of-the-art LRMs are guided by "normal" software, like agentic AI systems or Google DeepMind’s AlphaProof Nexus, researchers are also interested in understanding stand-alone reasoning models that rely solely on their self-generated reasoning traces.

Anthropic has published details of a practical key-recovery attack targeting HAWK-256. This development, related to AI CODE CREATION, highlights potential vulnerabilities in cryptographic systems. The information was shared within the developer community, indicating its relevance for those involved in enterprise-grade security and premium support, emphasizing the ongoing need for robust cryptographic solutions.

Noisegate is a differential-privacy gateway designed to provide AI agents with query access to sensitive data, ensuring that no individual's record can leak. It achieves this through a trusted privacy engine that clamps data to declared ranges, adds calibrated Laplace noise, and decrements a budget. The system includes a validation/guardrail layer for policy checks and handles untrusted LLM compiler output by emitting a CONSTRAINED QUERY AST, not free SQL, via structured/schema-forced output.

The OlmoEarth Platform, detailed at allenai.org/olmoearth, enables geospatial inference at a planetary scale. Its independent partitions allow stages to run across thousands of compute instances simultaneously. This was demonstrated by generating a North American wildfire-risk map, utilizing 19,600 CPUs and 994 GPUs in parallel, achieving a 155x speedup by reducing 4,737 hours of serial compute to 30.5 hours. The platform also aims to lower barriers to geospatial model use through agentic tools and interfaces.