July 17, 2026

The release addresses LLVM producing incorrect optimized code under particular conditions. Because a Rust 1.97.0 intermediate-representation change increased the likelihood, the team both backported the LLVM fix and disabled that Rust-side change. The underlying miscompilation predates 1.97.0 and is known to exist since at least Rust 1.87.

Users can move stable toolchains with rustup update stable. CI configurations that pin versions should explicitly select 1.97.1 and regenerate optimized artifacts. Preserve artifact hashes, critical numerical results, and tests across optimization levels so impact analysis covers emitted binaries rather than assuming that only projects briefly on 1.97.0 need review.

Temporal separates meanings into types such as PlainDate for calendar dates, PlainTime for wall-clock time, Instant for absolute time, and ZonedDateTime for a time-zone-aware value. Values are immutable and months are one-based, unlike Date. The specification is settled for planned ECMAScript 2027 inclusion, with unflagged support in current Chrome, Edge, Firefox, and Node.js versions cited by the guide.

Because Safari support remains in progress, public-web adoption still needs an explicit compatibility target. Existing Date values can interoperate through methods such as toTemporalInstant, enabling gradual replacement around UTC storage, time-zone conversion, or calendar arithmetic. Add daylight-saving transitions, nonexistent local times, and month-end behavior to tests, and measure both bundle cost and behavior if a polyfill is required.

Inkling uses a mixture-of-experts design with 975 billion total parameters and 41 billion active for inference. It supports up to one million tokens and was pretrained on 45 trillion tokens spanning text, images, audio, and video. Full weights are available, while the preview Inkling-Small variant activates 12 billion parameters.

Thinking Machines explicitly presents Inkling as a customizable foundation rather than the strongest model overall. Tinker supports fine-tuning, but adopters should separate training, evaluation, and holdout data, then measure modality-specific regressions, long-context cost, and reproducibility. Open weight access is valuable evidence of inspectability, not evidence that a deployment meets an application's quality or safety bar.

Grok Build centers on a mouse-aware full-screen TUI for repository understanding, file modification, command execution, web search, and long-running tasks. It also provides a headless mode for scripts and CI plus Agent Client Protocol support for editor embedding. Prebuilt binaries cover macOS, Linux, and Windows, with source-build instructions for Rust-based development.

The repository is periodically synchronized from xAI's monorepo, and SOURCE_REV records the corresponding internal commit, so the public tree may not be the development tip. Source-tree Windows builds are described as best effort and untested. Pin a commit and test writable paths, allowed commands, network egress, and failure stops before granting the agent production credentials.

Parallel Web Search is now a native grounding provider for the Gemini Enterprise Agent Platform through the Gemini API, Agent Studio, and Google Cloud Marketplace. It searches public web data and returns citations supporting the result. Teams may also extract, cache, or post-process the retrieved material with other models in a larger workflow.

Use requires a Marketplace subscription and acceptance of provider terms and pricing, with charges appearing on the Google Cloud invoice. Zero data retention is optional rather than an assumed organizational default. Benchmark citation agreement, stale-source handling, timeouts, no-search fallback, and total task cost, then review which prompts and sensitive data cross the provider boundary.

Advanced search extends field filters with AND and OR expressions. The reviews: qualifier uses the Reviewers field, making it easier to combine ownership, status, dates, and review queues in one view. With general availability, teams can treat these expressions as shared operating views rather than relying on personal, informal query strings.

GitHub is separately removing deployment-status records older than ninety days while preserving the current state for each environment. Any team using Projects as a long-term audit or trend source should export the required history through an API or another system. Inventory existing views and document search logic alongside retention requirements.

The %%bqsql cell magic executes SQL in BigQuery and can assign the result to a BigQuery DataFrame. Queries may reference BigQuery or Cloud Storage tables and temporarily uploaded local pandas DataFrames. BigQuery DataFrames' lazy execution keeps large joins and aggregations server-side instead of materializing an entire intermediate dataset on the notebook client.

Users load the BigQuery DataFrames extension and must provide a project ID even in the sandbox. Advanced services such as BigQuery ML can require billing, so templates should specify project, region, maximum bytes billed, and temporary-data cleanup. Port one existing analysis and compare results, scanned bytes, runtime, and client transfer before broader adoption.

The incident ran from 16:45 to 20:18 JST on July 16. The final service-health explanation cited an internal constraint in the fleet managing private VPC-origin connections, which prevented routing configuration from loading correctly. An existing Cloudflare Workers failover served static pages but could not restore the dynamic preview workflow, forcing a second recovery path.

Construction began at 18:03. Traffic moved from CloudFront to a public Northern Virginia ALB, across inter-region VPC peering, then through a fixed-address Tokyo NLB to the existing ALB and ECS service; normal behavior was confirmed at 18:42. The temporary path adds public-subnet, IPv4, and security tradeoffs and is slated for removal. Rehearse alternatives from minimum dynamic functionality and recovery objectives rather than copying this topology blindly.

According to Mindgard, Cursor's Windows Git discovery can select a git.exe at the workspace root and execute it multiple times merely when the repository opens. The firm says it first reported the issue on December 15, 2025, reproduced it against Cursor 3.2.16 on April 30, 2026, and published full details on July 14.

No Cursor response, CVE, or independent verification was presented with the disclosure, so scope and patch status should remain attributed to Mindgard. Interim controls include opening untrusted repositories only in Windows Sandbox or a VM, denying execution from workspace paths through AppLocker or Windows App Control, and monitoring unexpected git.exe launches with endpoint detection.