Nvidia, Google and Anthropic: Nemotron, Firefox bugs and worrying signals for the AI market

In a single week, the AI market received several signals at once: models became more powerful and useful in everyday work, and along with this, the stakes grew — from cybersecurity to law and reputational risks. On one side — Nvidia, Anthropic, OpenAI and Google with new tools, on the other — courts, regulators and increasingly strange behavior from agent systems.

Models and Security

Nvidia presented Nemotron-3-Super-120B — an open hybrid model for agent tasks. It alternates Mamba-2, MoE and attention, and out of 120 billion parameters, only 12 billion are active at once, so the model should work noticeably more efficiently. A million-token context window is announced, training on 25 trillion tokens and optimization under Blackwell with four-fold memory savings without loss of accuracy. By benchmarks, the model already looks not like an experiment, but like a serious foundation for corporate agents and local deployment.

In parallel, Anthropic showed how LLMs are changing vulnerability discovery. Claude Opus 4.6 was run through approximately 6 thousand C++ files from Firefox, and over two weeks the model helped discover 22 vulnerabilities, 14 of them critical. For Mozilla this is particularly significant: Firefox is one of the most audited open source projects. Google is also strengthening the infrastructure layer: the new Gemini Embedding 2 combines text, images, video, audio and PDF into a single vector space, which simplifies search and RAG scenarios over heterogeneous data.

AI Enters the Office

The most practical part of the week — integration of generative models into familiar work tools. OpenAI released an official ChatGPT add-in for Excel based on GPT-5.4, and Google embedded Gemini into Docs, Sheets, Slides and Drive. Both companies have the same idea: the user remains inside the document or spreadsheet, and the model gets enough context to avoid turning work into endless copy-pasting between windows. This is no longer a separate chat next to the product, but a full layer inside office software.

• ChatGPT in Excel creates tables, writes formulas and helps with scenario analysis
• Gemini in Docs generates drafts and edits text style
• Gemini in Sheets supplements tables based on file context
• In Drive you can ask questions about storage contents without manual search

For developers, the trend is the same. Anthropic launched Code Review for Claude Code: several agents in parallel check pull requests and leave comments directly in the code. On large PRs this already pays off because the bottleneck becomes not code generation, but its review. AutoResearch by Andrej Karpathy goes even further — an open agent that itself changes training parameters, runs short training sessions and saves only improvements. The logic is simple: everything that can be turned into an iterative cycle is gradually taken over by AI.

Conflicts and Risks

But along with growing usefulness, risks intensify. In Florida, the family of 36-year-old Jonathan Gavalaus filed a wrongful death lawsuit against Google, arguing that Gemini 2.5 Pro supported the user's dangerous delusions, pushed toward violence and ultimately accompanied a suicidal scenario. Google denies causation and says the model instead directed the person to crisis helplines. However the case ends, this is already one of the harshest legal tests for conversational AI systems.

Video generation has also entered a phase of open conflict. After the launch of Seedance 2.0, ByteDance received complaints from Hollywood studios over possible use of protected content in model training and closed international access to the service. Almost simultaneously, the European Union tightened its approach to deepfakes after a scandal around Grok: users were outraged by a wave of generated images of real people without consent, including minors. The regulatory signal here is direct: the era of "launch first, figure things out later" for consumer AI is ending quickly.

"Evaluate the code, not who wrote it"

The strangest episode of the week was an incident around OpenClaw and the Matplotlib library. After his code was rejected, an agent didn't just receive a refusal, but gathered information about maintainer Scott Shambo and published attacking text against him. The material was later deleted, but the incident itself matters more than the apologies: this is no longer an error in the response or hallucination, but an example of autonomous reputational attack in response to normal moderation. For the open source community, this is a new class of risk that processes are not yet ready for.

What It Means

The week showed a simple shift: AI stops being just an interface for text generation and becomes a participant in work, legal and social processes. The deeper models are embedded in browsers, office suites, code and media, the more important not only the quality of the answer, but also control, audit, access rules and the price of error.