- NATURAL 20
- Posts
- Gemini 4 Carbon Pushes AI Cybersecurity Into a New Era
Gemini 4 Carbon Pushes AI Cybersecurity Into a New Era
PLUS: NVIDIA's 40-language speech model, GitHub's autonomous projects, and a new Antigravity plugin marketplace.
Wes Roth
October 30, 2026
In partnership with
The AI notetaker that gets the hard words right
Your AI is only as good as what you feed it. Feed it a transcript with your product names, acronyms, and numbers spelled wrong, and you spend the afternoon hand-fixing the output.
Wispr Flow Notetaker uses your dictionary and calendar, so product names, acronyms, numbers, and uncommon names come out spelled right. It pops up when your meeting starts and captures Zoom, Google Meet, Teams, and Slack huddles with one click, no bot joining the call.
Then your meetings show up inside Claude or ChatGPT through the built-in connector. No copy paste. Try Notetaker free on Mac and Windows.
Today:
Google unveils Gemini 4 Carbon with stronger cybersecurity benchmarks
Mistral releases the open weights of its trillion-parameter Le Chonk model
OpenAI and Oracle plan to bring enterprise applications into ChatGPT
NVIDIA launches Nemotron 3.5 Transcribe for 40 languages
Google Antigravity opens a marketplace for agent plugins
AI Models Are Moving Into Security and Enterprise Work
Google has introduced Gemini 4 Carbon with a focus on cybersecurity and complex engineering, while Mistral has released the downloadable weights of its most capable model, Le Chonk. OpenAI is also working with Oracle to bring business data and workflows into ChatGPT.
The announcements show three directions for advanced AI: models that help find and fix security flaws, powerful systems organizations can operat
Google announced Gemini 4 Carbon, a new frontier model built for demanding cybersecurity, coding and long-running engineering tasks. The company says Carbon can persist through complicated work over multiple days, including investigating difficult bugs, testing vulnerabilities and generating fixes across large codebases.
In Google's published evaluations, Carbon scored 88.2% on SEC-bench Pro for vulnerability discovery, 54.8% on ExploitGym for validating exploitable weaknesses and 71.0% on CWE-bench for patching. It also ranked first on Artificial Analysis Intelligence Index v5.0 with a score of 59 when Google announced the model.
Google reports improvements in model behavior as well as capability. In internal coding tests, high-severity misalignment alerts declined by nearly 80% compared with Gemini 4 Argon. Carbon also recorded a 0% success rate for tested indirect prompt-injection attacks on Gray Swan's benchmark. These results are benchmark and internal-test measurements, not guarantees against real-world attacks.
The model is initially going to a small group of cybersecurity defenders and trusted partners through Google's Fairwind Program. Google is participating in voluntary U.S. government pre-release testing and plans broader developer, business and consumer access after further safeguards testing.
When Carbon becomes generally available, Google plans introductory API pricing of $2 per million input tokens and $10 per million output tokens, matching Argon's launch rates. Standard pricing after the promotional period is planned at $4 and $20 respectively.
Mistral released the weights of an improved Mistral Large 4, also called Le Chonk. The model was first introduced in public preview on October 6, but organizations can now obtain its weights to customize and run it on their own infrastructure.
Large 4 is a mixture-of-experts model with roughly one trillion total parameters and approximately 52 billion active parameters including associated model components. It supports text and image inputs and a context window of up to one million tokens, targeting complex software projects, business workflows and security operations.
Mistral says the updated release ranks first among open-weight models on Cognition's FrontierCode benchmark and has improved its performance on coding and defensive cybersecurity tasks. Testing and post-training work with government organizations and corporate security partners informed the release, with results reported by Mistral and independent benchmarking groups.
The open-weight release gives organizations more control over deployment and data handling than a hosted-only model. Teams can fine-tune it with Mistral Forge or host it using supported inference frameworks, while the license places specific conditions on model-as-a-service offerings.
The model's scale also means self-hosting requires substantial computing hardware. Its release creates another option for companies that want advanced AI capability without sending every request to an external model provider.
OpenAI and Oracle announced an expanded partnership to connect Oracle applications and business data with ChatGPT. The effort centers on Oracle AI Workspace, an experience under development that would let employees find information and complete approved tasks across finance, procurement and planning systems from one interface.
Instead of opening several business applications, an employee could ask about purchasing decisions, investigate expenses or check what remains before month-end financial close. The proposed integrations are designed to respect an organization's existing access permissions and approval processes.
Oracle also unveiled related offerings built with OpenAI technologies. Oracle AI Security Mesh targets vulnerability discovery and remediation using cyber models; Oracle App Composer is designed to help teams build and deploy applications linked to authorized company data; and Oracle AI SMB Bundle combines ChatGPT Business, Codex and Oracle infrastructure resources.
OpenAI said Oracle has more than 130,000 active ChatGPT users and over 95,000 active Codex users internally. Oracle reported that work requiring two to four days of recruiting market research can now be prepared in about 15 to 20 minutes with ChatGPT Work.
U.S. Oracle customers can purchase access to the OpenAI API, Codex and ChatGPT Work through Oracle Marketplace using eligible procurement arrangements and Oracle Universal Credits. Oracle AI Workspace and several newly announced products describe planned capabilities rather than fully completed integrations.
🧠RESEARCH
ChronoReplay trains video-language models on sequences of spatial tasks instead of isolated question formats. It combines different kinds of position and distance supervision while preserving their chronological order. The researchers report improvements across three benchmarks, including a 13.41-point gain on OST-Bench, helping models transfer spatial knowledge between different questions.
EmbeddingGemma 2 introduces an open model that converts text, images, audio and video into a shared representation for search and retrieval. The researchers report strong performance against specialist models and larger alternatives. Its adjustable embedding dimensions, optional modality components and quantized weights are designed to support deployment with limited computing resources.
Nemotron-Labs-Diffusion-Uni is a discrete diffusion language model that handles both visual understanding and image generation. Its training transfers abilities from vision-language and image-generation models, while a progressive refinement process improves generated outputs. Researchers report leading results among comparable diffusion models on visual reasoning, text-to-image generation and editing tasks.
📲SOCIAL MEDIA
🗞️MORE NEWS
NVIDIA introduced Nemotron 3.5 Transcribe, an open-weight speech model with about 30 billion total parameters and 4.3 billion active parameters. It can transcribe speech in 40 languages, identify speakers, translate recordings, follow transcription-style instructions and return word-level timestamps. NVIDIA also released checkpoints and fine-tuning tools for developers building meeting assistants and other speech applications.
Google launched a Plugin Marketplace for Antigravity's app and command-line tools, giving developers a simpler way to install integrations without configuring each tool separately. Plugins can bundle skills, external API connections, persistent rules, execution hooks and specialized agents. Initial partners span project management, design, security and data services, including Atlassian, Canva, Datadog, Figma and Stripe.
GitHub introduced Projects in the Copilot app, letting users give agents an objective with context from repositories, designs and connected tools, then review the resulting work. Copilot also added computer use in public preview on macOS and Windows, allowing approved interaction with desktop applications that may not provide APIs. The changes extend Copilot beyond writing code into coordinating broader workflows.
Perplexity announced that its enterprise offering achieved FedRAMP Moderate certification, expanding its government-focused service with live web research and source-linked answers inside a security-controlled environment. It says government queries stay within its own search infrastructure and that administrative controls support identity, data retention and audit requirements. The announcement follows Perplexity's earlier federal certification at a lower security classification.
Anthropic detailed a self-assessment system in which Claude evaluates conversations after responding and flags potentially harmful use for offline review. In testing with Claude Fable 5, it identified 78% of conversations and 94% of high-priority cases flagged by a stronger reference classifier at a 1% false-positive rate, while adding less than 2% estimated inference compute. The technique has been deployed since June; Anthropic published the underlying research on October 16 and emphasizes that it supplements rather than replaces other safeguards.
What'd you think of today's edition? |
Reply