Retro cartoon cover showing a microphone conversation with an AI voice console, simultaneous sound waves, delegation to a larger machine, and an API plug panel.

OpenAI GPT-Live Leads July 9 Tech Briefs

OpenAI's full-duplex GPT-Live voice model leads an edition covering Linux kernel flaws, post-quantum signatures, AI developer tooling, cloud concentration, security incidents, and open-source systems.

Mike Chumba Mike Chumba
9 min read
1795 words

OpenAI launched GPT-Live, a voice model family for ChatGPT Voice built around a full-duplex architecture that can listen and speak at the same time.

OpenAI says GPT-Live can hand harder tasks to GPT-5.5 while keeping a voice conversation active. The company says GPT-Live-1 and GPT-Live-1 mini are rolling out in ChatGPT Voice, with API availability planned later.

Featured source: OpenAI , The Register .

Januscape KVM flaw and GhostLock Linux bug surface

Ars Technica reported that two high-severity Linux kernel vulnerabilities surfaced this week: CVE-2026-53359, Januscape, a KVM/x86 guest-to-host escape class bug, and CVE-2026-43499, a local privilege-escalation flaw in the rtmutex/futex path. The Januscape disclosure says the bug can be triggered from a guest on Intel and AMD KVM hosts and was patched in mainline, while NVD lists CVE-2026-43499 with a kernel.org CVSS 7.8 score and fixed stable versions.

Filed from: Ars Technica , oss-sec , NVD .

Cloudflare backs ML-DSA for the first post-quantum signature migration

Cloudflare published a post arguing that ML-DSA should be used for the first post-quantum signature migration instead of waiting for newer signature schemes now moving through NIST review. Cloudflare says most traffic it handles is already using ML-KEM encryption, while authentication still needs post-quantum signatures; NIST’s digital-signature project lists additional candidates under evaluation after the earlier post-quantum standards work.

Filed from: Cloudflare , NIST .

GitHub Agentic Workflows automate cross-repo documentation

GitHub published an Aspire team case study showing GitHub Agentic Workflows turning merged product changes into reviewed documentation pull requests across repositories. The post says Aspire 13.3 and 13.4 produced 82 merged feature-doc pull requests at a median of 44.8 hours after the product pull request, with engineer review; GitHub’s Agentic Workflows docs describe the system as public preview automation in GitHub Actions with scoped write handling.

Filed from: GitHub Blog , GitHub Agentic Workflows docs .

Ollama raises $65 million and grows to nearly 9 million users

TechCrunch reported that Ollama raised a $65 million round led by Benchmark and has grown to nearly 9 million users. The report says the open-source local AI tool has about 176,000 GitHub stars and nearly 17,000 forks, and Ollama’s own site now presents a cloud tier alongside its local open-model runtime.

Filed from: TechCrunch , GitHub , Ollama .

Entire launches a distributed Git hosting network for coding agents

Entire, a company founded by former GitHub CEO Thomas Dohmke, launched a Git-compatible hosting network that mirrors GitHub repositories for agent-heavy clone and fetch workloads. Entire says the first rollout covers existing users and regions in the US, EU, and Australia, with new users waitlisted; the company also says it plans to open source the Git network and add self-hosting.

Filed from: The Register , Entire .

Microsoft patches RoguePlanet Defender zero-day

The Register reported that Microsoft shipped a fix for CVE-2026-50656, the RoguePlanet elevation-of-privilege flaw in the Microsoft Malware Protection Engine used by Defender. NVD’s July 8 change history lists the affected Microsoft Malware Protection Engine versions as earlier than 1.1.26060.3008, and records the issue as a high-severity local privilege-escalation vulnerability with public proof-of-concept code.

Filed from: The Register , NVD .

SambaNova previews SN50 with H200 GPUs for MiniMax inference

SambaNova said its RAISE Summit 2026 SN50 preview ran MiniMax M2.7 in a heterogeneous setup with four Nvidia H200 GPUs for prefill and 16 SN50 RDU chips for decode. The company says Artificial Analysis benchmarked the demo at up to 850 tokens per second on short-context workloads and over 450 tokens per second on long-context workloads; The Register separately reported a 763 tok/s result for MiniMax M2.7 in third-party testing.

Filed from: SambaNova , The Register .

Microsoft moves Commercial Cloud currency pricing to annual January updates

The Register reported that Microsoft is moving Commercial Cloud local-currency pricing updates from semiannual revisions to annual updates every January. Microsoft’s Partner Center announcement says the FY27 transition begins July 1, 2026, the next local-currency update will take effect January 1, 2027, and guidance for each upcoming year will be issued every November.

Filed from: The Register , Microsoft Learn .

UK cloud outage report estimates billion-pound regional exposure

The Register reported that the Cyber Monitoring Centre and Parametrix estimated a 24-hour outage in major AWS regions could cause direct UK revenue losses of about GBP1 billion for eu-west-1 or GBP650 million for us-east-1. The report says more than 60 percent of UK companies depend on cloud services for critical functions, rising above 80 percent among FTSE 100 firms.

Filed from: The Register , Computer Weekly .

Accenture confirms isolated matter after alleged 35GB data listing

The Register reported that Accenture confirmed an isolated security matter after a cybercriminal claimed to be selling 35GB of internal material, including source code, keys, tokens, and configuration files. SecurityWeek separately reported that Accenture said it remediated the source and saw no impact to operations or service delivery, while the company did not confirm the volume or full contents of the claimed data.

Filed from: The Register , SecurityWeek .

QuantumDiamonds adds funding after EU-backed semiconductor grant

TechCrunch reported that Munich startup QuantumDiamonds raised a EUR15 million equity round alongside a previously approved EUR76 million German state-aid package for semiconductor testing equipment. The European Commission said the aid supports a Munich facility for advanced semiconductor metrology and inspection systems based on quantum sensors, with the project treated as first-of-a-kind in Europe.

Filed from: TechCrunch , European Commission .

General Intuition trains physical-AI models from game action data

TechCrunch reported that General Intuition trained a model on millions of hours of video-game data and demonstrated it controlling a quadruped robot after fine-tuning on eight minutes of real-world robotics data. General Intuition’s own site says it raised a $320 million Series A to build models that perceive, predict, and act in virtual and physical environments, and says its models learn from action-labeled video datasets.

Filed from: TechCrunch , General Intuition .

Microsoft plans to retire OWA Light in Exchange Server

Microsoft announced that it plans to retire and disable the OWA Light experience in Exchange Server in a future update. BleepingComputer reported that Microsoft expects the removal in an August 2026 Exchange Server update; Microsoft says users should move to the modern Outlook on the web experience, and admins can already block OWA Light through mailbox-policy and virtual-directory settings.

Filed from: Microsoft Community Hub , BleepingComputer .

Temasek targets up to 15 percent AI portfolio exposure by 2031

The Register reported that Singapore sovereign wealth fund Temasek plans to increase AI exposure from about 6 percent of its portfolio to up to 15 percent by 2031. Temasek’s annual-review speaker notes identify AI and core-plus infrastructure as target areas, citing AI, electrification, data-centre growth, and energy-transition investments; Temasek also reported net portfolio value of S$518 billion for the year.

Filed from: The Register , Temasek speaker notes , Temasek annual review release .

Meta is reportedly developing always-aware smart glasses

The Verge reported, citing the Financial Times, that Meta is working on prototype “super sensing” smart glasses that could continuously record audio and take photos every few seconds for later AI queries. Meta told The Verge it does not comment on internal prototypes and said it is working on glasses with privacy built in; The Verge also noted Meta announced a separate update that disables the camera if the recording LED is tampered with.

Filed from: The Verge , Financial Times .

Google SynthID watermark helped identify a political hoax image

TechCrunch reported that Google’s SynthID watermark system was used to identify a viral AI-generated image falsely depicting Senator Mitch McConnell in a hospital bed. The report says Snopes found the image registered as containing a SynthID watermark, and TechCrunch notes that SynthID only works when an image-generation tool participates in the program.

Filed from: TechCrunch , SynthID .

New Horizons wakes after a 321-day hibernation

The Register reported that NASA’s New Horizons spacecraft woke after 321 days in hibernation and resumed operations beyond Pluto. NASA says flight controllers confirmed the wake-up on June 23, with the spacecraft about 5.9 billion miles from Earth, and that the team will begin downlinking health data followed by data from three scientific instruments.

Filed from: The Register , NASA .

From the Community

Mistral releases Robostral Navigate for single-camera robot navigation

Mistral AI released Robostral Navigate, an 8B-parameter model that enables robots to navigate environments using only a single RGB camera, without depth sensors or LiDAR, achieving 76.6 percent success on the R2R-CE benchmark for unseen environments. The model was trained in simulation using prefix-caching and online reinforcement learning, and Mistral says it generalizes across wheeled, legged, and flying robots.

Filed from: Mistral AI .

SpaceWASM provides a NASA/JPL WebAssembly interpreter for spacecraft sequencing

NASA JPL open-sourced SpaceWASM, a flight-compliant WebAssembly 1.0 interpreter for safety-critical spacecraft execution, featuring a streaming decoder and deterministic memory allocation. The interpreter converts WASM bytecode into an intermediate representation optimized for low peak memory and is written in Rust.

Filed from: GitHub .

Chatto releases an open-source self-hosted group chat application

Chatto, a group and team chat application, has been released as open source under version 0.4, offering self-hosted binaries for Linux, macOS, and Windows with end-to-end encrypted voice and video calls. The developer also announced Chatto Cloud, a paid hosting service with European infrastructure entering public beta soon.

Filed from: Blog post .

pgrust rewrites PostgreSQL in Rust and passes regression tests

A project called pgrust has rewritten PostgreSQL in Rust and now passes all 46,000-plus regression tests for Postgres 18.3. The implementation is disk-compatible with Postgres and can boot from an existing Postgres 18.3 data directory, while the project says existing Postgres extensions and procedural languages are not generally compatible.

Filed from: GitHub , Launch blog post .

DocuBrowser indexes local documents into an offline knowledge base

DocuBrowser is an open-source tool that indexes local documents, including PDFs, ebooks, and Word files, into a searchable knowledge base using SQLite FTS5 keyword search and semantic similarity through local Ollama models. It runs offline and supports hybrid search, AI-generated synopses, PII detection, and multiple document formats.

Filed from: GitHub .

Developers migrate projects from GitHub to Codeberg and self-hosted forges

How-To Geek reported that Ghostty, Zig, Tenacity, and other open-source projects have left GitHub for Codeberg, SourceHut, or self-hosted forges such as Forgejo. The article cites maintainers discussing GitHub outages, Microsoft’s AI integration with Copilot, and political concerns around GitHub’s contract with ICE.

Filed from: How-To Geek .

llmpl adds OpenAI-compatible LLM calls to SWI-Prolog

llmpl is a minimal SWI-Prolog library that exposes llm/2 to post prompts to any OpenAI-compatible chat/completions endpoint and unify the response. The library supports configurable endpoint and model settings, plus a reverse mode that asks the model to suggest a prompt for a desired response.

Filed from: GitHub .