top of page

Voice of AI: New chips, new models, new Office and a button with consequences ⚡

14 hours ago
5 min read

A single click on "Don't ask again" has unleashed around 2,500 passwords and API keys onto GitHub. That's the smallest piece of news this week, but perhaps the most instructive. Meanwhile, Vera Rubin has arrived in the cloud, Claude Fable 5.2 is getting faster and cheaper, and Anthropic is challenging Microsoft and Google on their own turf with Sheets and Whiteboards . Who actually controls what your AI agents approve on behalf of your employees?



🔑 One click, 2,539 open doors

The security firm Apiiro has found 2,539 instances of vulnerability in 1,994 public GitHub repositories. All can be traced back to the same source. Claude Code prompts for permission before every shell command. Users who select "Yes, and don't ask again" save the command verbatim as a rule in the file `.claude/settings.local.json`. If the command contained a password, it is now stored in plain text in a project file, which is often included in the next commit. Anthropic is investigating an automatic cleanup and recommends using auto mode until then.

The real lesson: The model didn't fail here, but rather the governance. Where no one sets rules, convenience wins out over security. This is precisely where we at OAK AI come in. Our AI Readiness Check not only assesses whether a company can deploy AI, but also whether it securely manages AI agents. What you can do immediately: run `git ls-files .claude/settings.local.json` in every repository, replace the affected keys, and add the file to your `.gitignore` folder.


⚙️ Vera Rubin: Hardware for AI agents

Since October 27, NVIDIA's VeraRubin NVL72 systems and the Vera CPU have been generally available in Oracle Cloud. Vera is NVIDIA's first fully in-house designed processor. It is not designed for pure computing power, but rather for the work of AI agents: calling tools, executing code, and handling many tasks in parallel. Oracle plans to deploy multi-gigawatt data center clusters.

It's noteworthy that Oracle also offers Google TPUs as well as chips from SambaNova and Cerebras. For businesses, this means genuine choice when it comes to computing power. NVIDIA estimates the revenue potential at up to $40 billion per gigawatt. This is a manufacturer's figure, but it illustrates the scale the industry is now planning on.


🧠 Fable 5.2: faster, cheaper, but check it out yourself

Claude Fable 5.2 has been generally available since November 9th. According to Anthropic, the model is up to 30 percent faster, and the price has dropped by 20 percent to $8 or $40 per million tokens. The even more powerful Mythos 5.2 remains reserved for certified life sciences and cybersecurity organizations.

Clarity instead of hype: The performance figures come from the manufacturer or from internal tests by individual customers; they have not been independently verified. A different finding is more meaningful. METR, an independent evaluator with direct access to Anthropic, estimates that the company is now progressing about twice as fast as before on clearly defined tasks, primarily because it uses AI internally. Model cycles are therefore becoming even shorter. Anyone who commits to a single provider today will pay for it tomorrow with switching costs.


📊 Claude Sheets, Whiteboards, Data: Attack on the Office

On November 10th, Anthropic launched Claude Sheets and Claude Whiteboards in beta. The existing dashboards are now generally available under the name Claude Data, and Claude Meetings has been announced. This makes Anthropic a direct competitor of Google and Microsoft, both of which have stakes in Anthropic. Anthropic describes Sheets as "fairly basic," meaning rather simple. For financial models, the company continues to refer users to its extensions for existing spreadsheet programs.

Important for IT managers: In enterprise plans, spreadsheets and whiteboards remain disabled until an administrator activates them. Claude Data, on the other hand, will be automatically enabled on November 19th. Anyone who doesn't review their data sharing settings by then will have a new tool with access to company data starting Thursday, a tool that no one has authorized.


⚡ Electricity, capital, minds: What it all stands for

Three news items this week show the same pattern. According to the Financial Times, Anthropic paid approximately $3.4 billion for a delivery site for 1.7 gigawatts of gas generators. The stock is scheduled to begin trading on the Nasdaq on November 19 under the ticker symbol ANT, with a price range of $70 to $78 and an issue volume of up to $64.4 billion. Meanwhile, at NVIDIA, CEO Jensen Huang resigned on November 12. This followed an investigation into an affair, with an unproven suspicion of espionage lurking in the background. CFO Colette Kress is leading the company on an interim basis; the quarterly results will be released on November 17.


Technology is advancing, and so are dependencies. Anyone evaluating AI providers, for example with a technology due diligence like our ARGUS, must therefore examine not only model quality and price, but also energy supply, capital structure, and dependence on individual people.


📌 Key messages

  • A convenience button in Claude Code has exposed 2,539 login credentials: Rules for AI agents belong in governance, not in habit.

  • Vera Rubin and the Vera CPU are live: The hardware is being tailored to AI agents, and for the first time there is a real choice between chip suppliers.

  • Fable 5.2 is faster and 20 percent cheaper, but the performance figures have not been independently verified.

  • Claude Data will be automatically enabled in Enterprise plans on November 19th, so please check your admin settings now.

  • Energy, capital, and individuals are the new risk factors for AI infrastructure.


📚 Sources


Conclusion

The most important insight this week isn't found in a chip or a model, but in a dialog box. AI agents do exactly what we allow them to do, even when we haven't consciously decided otherwise. Companies that establish clear rules now for permissions, data access, and the choice of their providers are harnessing the pace of the next generation of AI models. Everyone else is playing catch-up.

Want to know how securely your organization manages AI agents today? Write to us at info@oakai.de or start the AI Readiness Check at readiness.oakai.de .


The future is not a matter of chance. It is a choice. Ralph Schwehr, OAK AI



 
 

OAKAI®

NEWSLETTER

Sign up for the OAKAI newsletter.

  • LinkedIn

© 2024 OAKAI®

Imprint

Data protection

Thanks for subscribing!

bottom of page