Gemini 4 Carbon now broadly usable, new agent backdoors, check Notion chats for leaks

Important AI news

Microsoft describes new backdoor technique for AI agents

Microsoft Research warns about “Lazarus”. A few targeted, otherwise inactive neurons are enough to train a model on a hidden trigger. A backdoor here is a secret trigger in the model. It starts harmful actions on a signal without being noticed. This is especially risky for agents with tool or terminal access. Protection depends heavily on model provenance according to Microsoft. That means a verifiable origin and integrity record of weights across the supply chain. Foundry checks provenance, ties models to trusted publishers, and plans signatures with verifiable lineage. For you this means: only grant agents the access they really need, watch their actions, and treat model weights like production code. The results and mitigations come from the vendor. An independent review is not available.¹

IBM and Microsoft announce “Imagine Enterprise”

Both companies present a platform for industry agents. IBM Consulting brings over 6,000 prebuilt agents and assets by its own account. Microsoft supplies AI, cloud, and security tech. The goal is regulated deployments with governance, observability, and AI compliance, first for financial services, insurance, and retail. It is an announcement with accompanying engineering support for qualified projects. Prices are missing. If you start in a regulated environment, you can build faster from templates instead of starting at zero. Figures and capabilities come from the press release.²

Gemini 4 Carbon becomes generally available, prices rise on December 1

Gemini 4 Carbon is now more widely accessible. The launch price is 2 US dollars per million input tokens and 10 US dollars per million output tokens. On December 1 Google doubles prices to 4 and 20 US dollars. If you want to test Carbon, plan your evaluations before the price increase. Availability and prices come from Google and an independent blog post. External confirmations for performance claims are missing.³⁴

Notion AI: users report possible exposure of chat sessions

A user reports an email from Notion about the incident at provider Fireworks AI. Anonymous Notion AI chats in under 1% of workspaces could be affected. The poster writes that their download contains a complete chat session. Notion has disabled the Insights feature and stopped using Fireworks for open models. If you shared credentials in Notion chats, rotate them as a precaution and review the provided data. There is no official detailed analysis yet. The information comes from a user report.⁵

Highlights for your workday

• Google and Adobe are moving closer. You can edit PDFs directly in Drive and use Adobe features in Gemini. If your team works a lot with PDFs, you avoid tool switching. Check if the feature is already visible in your workspace.⁶

• OpenAI introduces “Businesses in ChatGPT”. You maintain your company info there, get insights into presentation, and route queries into actions. This ensures ChatGPT uses correct data about your business.⁷

• Meta presents Muse Ember, a small, efficient model for routines like tool calls, extraction, and long documents. Prices: 0.10 US dollars per million input and 0.50 US dollars per million output up to 200k tokens, above that more expensive. The vendor points to good efficiency in third-party charts. Independent evaluations vary. For you this means cost-effective automation for standard tasks.⁸

Tools and updates

• Mitra-v2, a tabular foundation model under Apache-2.0, targets classification and regression on business data. The report says it leads on TabArena and provides prediction intervals. Use case: quick forecasts without a full ML restart. Requirement: Python setup and some ML know-how. Vendor information.⁹
• Aircad brings professional CAD to the browser, with NURBS modeling, DWG/STEP/3DM support, and an integrated AI agent. Use it for collaborative designs without local installation. Requirement: a current browser, data export depending on format.¹⁰
• NVIDIA donates OpenShell to the CNCF. The open source runtime helps control and verify agent activities deterministically. Useful if you want to run agents on Kubernetes securely. Requirement: Kubernetes expertise.¹¹
• Naive-N0.5-Flash targets extremely fast code generation. A test compares it with DeepSeek and GLM. Very high throughputs, sometimes longer reasoning times on hard tasks. Good if response time matters more to you than cost stability. Benchmarks are partly vendor, partly third-party measurements with caveats.¹²

Try it in five minutes

Own practice idea: find sensitive data in your AI chats and shut down the riskiest right away.

1) Collect three chat transcripts from an AI tool you use for work. Copy them into a local text file. For testing, use this sample text instead: “Our API_KEY=sk-test123 and the webhook https://hooks.slack.com/services/T000/B000/XYZ belong in the vault, not in chats.” 2) Search the file for secret patterns: “key=”, “token”, “secret”, “-----BEGIN”, “hooks.slack.com”, “Authorization: Bearer”, “ssh-rsa”, “AKIA”. Mark hits. 3) Result and check: list every finding and rotate the real credentials today. Check criterion: for every finding there is either a rotation or a documented confirmation that it was only test data.

Expected result: you know the leaks, have replaced real keys, and created a short do-not-paste list for your team. Context: the reported Notion incident shows that chats can be exposed unintentionally. Microsoft’s backdoor finding shows why strict privileges and clean key management matter.⁵¹

A good find

Designing for Agents with Zeno Rocha (Resend). Three practical points: make pricing pages machine-readable. For example with a .md variant that returns a table in one request. Mirror the key dashboard actions in your MCP server so agents can work without a UI. Open the returned tool data in your own sessions and trim it if it blows the context. Application for you: add a simple agent-readable pricing table today and test with your preferred model whether it extracts prices correctly.¹³

Sources

1. Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents (microsoft.com)
2. IBM and Microsoft announce a new AI agent platform designed for industry-specific transformation (news.microsoft.com)
3. Gemini 4 Carbon is now available (deepmind.google)
4. llm-gemini 0.36 (simonwillison.net)
5. Tell HN: Notion AI chat sessions have been exposed via Fireworks AI hack (news.ycombinator.com)
6. We’re bringing Adobe’s creative and productivity tools closer to your everyday Google apps. (blog.google)
7. Introducing Businesses in ChatGPT (openai.com)
8. Introducing Our Smallest, Most Efficient Model: Muse Ember (simonwillison.net)
9. Mitra-v2: Scaling synthetic pretraining for tabular foundation models (amazon.science)
10. Show HN: Aircad – Professional drafting and 3D modeling in the browser (news.ycombinator.com)
11. NVIDIA Expands Open Source Commitment With OpenShell Donation and New AI Infrastructure (blogs.nvidia.com)
12. NaiveAI Promises 2,100 Tokens per Second. How Well Does It Code? (latent.space)
13. Designing for Agents — Zeno Rocha, Resend (latent.space)