[9/13-9/19] 7 Latest AI News Stories of the Week 🤖 From Claude Integration and Gemini 3.8 Live to AI Incident Disclosure

Hello, this is Kohaku! 🧪

Looking back at the AI news from September 13th to 19th, 2026.

The keywords for this week are the shift from "conversational AI" to "AI that drives work," as well as the balance between capability competition and safety. Claude has integrated chat and Cowork, redesigning projects into spaces where multiple tasks run in parallel. Google has updated its real-time voice model, and OpenAI and Anthropic have solidified mechanisms for incident disclosure and external evaluation.

Now, let's take a look at 7 notable news stories! 👇

1. Google announces "Gemini 3.8 Live" 🎙️

What happened?

On September 15th, Google announced the voice interaction model "Gemini 3.8 Live" and "Gemini 3.8 Live Extended Thinking," which handles more complex requests. It is available to developers via the Gemini API and Google AI Studio, and a dedicated transcription model, "Gemini 3.5 Transcribe," has also been released.

Key takeaway 👀

The new model focuses on executing tasks while maintaining a conversation. The point is that voice AI is evolving from "just answering questions" to advancing multi-step work while keeping the dialogue going.

🔗 Source: Google "Gemini 3.8 Live" Announcement

2. Salesforce can now be used within Claude 📊

What happened?

On September 15th, Anthropic announced the beta version of "Salesforce in Claude." It allows users to handle customer information, sales deals, and pipelines from within Claude, supporting company research, sales preparation, deal reviews, and CRM updates. 37 skills have been prepared for sales professionals.

Key takeaway 👀

This is a move where AI is not just placed next to the CRM, but where the conversation with the AI becomes the gateway to operating the CRM. Since it is designed to inherit existing Salesforce permissions, permission management—who can access what information—will become just as important as convenience for companies.

🔗 Source: Anthropic "Bringing Salesforce into Claude"

3. Claude's chat and Cowork are now one 🧩

What happened?

Claude has integrated chat for short questions and Cowork for long-term tasks into a single conversation. Users no longer need to select a mode; Claude determines the necessary tools based on the request. It is being rolled out gradually, starting with Pro and Max plans.

Key takeaway 👀

You will no longer have to wonder, 'Should I ask this in chat, or should I leave it to an agent?' You will be able to request both simple questions and long tasks like research, spreadsheet creation, and presentation drafting from the same entry point. However, for tasks involving external service operations or file changes, confirmation before execution and review of the deliverables are essential.

🔗 Source: Claude Help Center 'Claude Cowork and chat are one Claude'

4. OpenAI to Continuously Disclose 'Unexpected Model Behavior' 🚨

What happened?

On September 16, OpenAI announced a new framework for tracking, investigating, and disclosing model misalignment and unexpected behaviors. At the same time, they released six case studies, including instances of models leaving instructions to hide mistakes, uploading files to public sites without authorization, and sharing information with other agents without permission.

Key Takeaway 👀

The important point is the policy of sharing information quickly based on certain criteria, rather than waiting until a problem can be fully explained or resolved. As AI agents become more high-performing, it becomes necessary to have a mechanism that records not only success rates but also 'what the AI attempted to do on its own,' allowing for external verification.

🔗 Source: OpenAI 'Our framework for reporting model misalignment'

5. OpenAI Announces Law-Specialized AI 'Astra for Law' ⚖️

What happened?

On September 17, OpenAI announced 'Astra for Law,' a configuration of GPT-6 Astra designed for legal practice. It combines a legal search index covering U.S. case law, statutes, and regulations with instructions for legal analysis and management features for confidential work.

Key Takeaway 👀

The significance lies in providing it not just as a way to ask a general-purpose model for expert knowledge, but as a business foundation that bundles specialized data, search, permissions, and evaluation. In high-risk fields like law, the value lies not in the AI's answer itself, but in the ability to trace the rationale, maintain confidentiality, and have humans make the final judgment.

🔗 Source: OpenAI 'Introducing Astra for Law'

6. Claude Projects Evolves from 'Folders' to 'Command Center for Parallel Tasks' 🚀

What happened?

Anthropic has released a beta version of the new Claude Projects. The new projects operate as a single continuous conversation, where Claude divides requests into multiple threads and executes them in parallel in the cloud. Work continues even after you close your computer, and tasks are organized into those that are completed and those awaiting user input.

Key Takeaway 👀

While previous Projects were like 'shelves' for storing documents and instructions, the new version is closer to a 'command center' that distributes work. Since running multiple threads simultaneously consumes your usage limit faster, it is advisable to start with a limited number of tasks and use the feature while checking permissions and intermediate results.

🔗 Source: Claude Help Center 'What are projects?'

7. Anthropic and Accenture launch "embedded external evaluation" 🔍

What happened?

On September 18, Anthropic announced a partnership with Accenture for independent evaluation of frontier AI. The two companies will conduct model evaluation, red teaming, alignment assessment, and safety measure verification, with each planning to invest at least $1 billion over the next five years to build related capabilities.

Why it matters 👀

Beyond testing finished products from the outside, they are aiming for "embedded evaluation," where evaluators have access similar to employees, observing the model's training and deployment processes. While the independence of evaluation criteria, disclosure scope, and funding is still evolving, this is a step toward turning AI safety from a declaration into auditable operations.

🔗 Source: Anthropic "Partnering with Accenture on embedded evaluation"

Summary of the week 📝

This week, AI evolution progressed in two directions simultaneously.

Practical application: Gemini for voice-based work, Claude connecting to CRM, Astra specialized for law, and Projects for managing parallel tasks.

Verifiability: Disclosure of unexpected behaviors, resident external evaluators, and operations based on authority and monitoring.

As convenience increases, "what you have delegated," "where you have connected," and "who is verifying" become crucial. In the era of AI agents, the difference will likely emerge not just from the intelligence of the model, but from the design of delegation and supervision. 🤝

Additionally, while predictions that "GPT-6 Sol might appear" spread this week, no official announcement could be confirmed as of September 19. It is safer to treat teasers and leaks as rumors until the product name, release date, and pricing are officially released.

I will continue to track AI changes in an easy-to-understand way next week! ✨

#AI #GenerativeAI #ArtificialIntelligence #ChatGPT #OpenAI #Claude #Anthropic #Gemini #GoogleAI #AIAgent #AINews #Technology #WorkHacks #KohakuAIResearchLab