Nvidia, Google and Anthropic: Nemotron, Firefox Bugs and Warning Signs for the AI Market

In a single week the AI market sent several signals at once: models got more powerful and more useful in everyday work, and the stakes rose with them, from cybersecurity to law and reputational risk. On one side are Nvidia, Anthropic, OpenAI and Google with new tools; on the other, courts, regulators and increasingly strange behavior from agentic systems.

Models and security

Nvidia introduced Nemotron-3-Super-120B, an open hybrid model for agentic tasks. It alternates Mamba-2, MoE and attention layers, and of its 120 billion parameters only 12 billion are active at any one time, so the model should run noticeably more efficiently. Nvidia claims a one-million-token context window, training on 25 trillion tokens and optimization for Blackwell with a fourfold memory saving without loss of accuracy. On benchmarks, the model no longer looks like an experiment but like a serious foundation for enterprise agents and local deployment.

At the same time, Anthropic showed how LLMs are changing vulnerability hunting. Claude Opus 4.6 was run over roughly 6,000 C++ files in Firefox, and in two weeks the model helped uncover 22 vulnerabilities, 14 of them critical. For Mozilla this is especially telling: Firefox is one of the most heavily scrutinized open source projects.

Google is also strengthening the infrastructure layer: the new Gemini Embedding 2 combines text, images, video, audio and PDFs in a single vector space, which simplifies search and RAG scenarios over heterogeneous data.

AI moves into the office

The most practical part of the week was the integration of generative models into familiar work tools. OpenAI released an official ChatGPT add-in for Excel based on GPT-5.4, and Google built Gemini into Docs, Sheets, Slides and Drive. Both companies have the same idea: the user stays 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 writing style
• Gemini in Sheets fills in tables based on the file's context
• In Drive, you can ask questions about the contents of your storage without searching manually

The trend is the same for developers. Anthropic launched Code Review for Claude Code: several agents review a pull request in parallel and leave comments directly in the code. On large PRs this already pays off, because the bottleneck is no longer generating code but reviewing it.

Andrej Karpathy's AutoResearch goes further still: an open agent that changes training parameters on its own, runs short training sessions and keeps only the improvements. The logic is simple: anything that can be turned into an iterative loop, AI gradually takes over.

Conflicts and risks

But as usefulness grows, so do the risks. In Florida, the family of 36-year-old Jonathan Gavalas filed a wrongful death lawsuit against Google, alleging that Gemini 2.5 Pro reinforced the user's dangerous delusions, nudged him toward violence and ultimately accompanied a suicidal scenario. Google rejects any causal link and says the model, on the contrary, directed the person to crisis helplines. However the case ends, it is already one of the toughest 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 the possible use of protected content in training the model, and it closed international access to the service. Almost simultaneously, the European Union tightened its approach to deepfakes after the Grok scandal: users were outraged by a wave of generated images of real people made without their consent, including minors. The regulatory signal is direct: the "launch first, sort it out later" era for consumer AI is ending quickly.

"Judge the code, not the person who wrote it"

The strangest episode of the week was the incident involving OpenClaw and the Matplotlib library. After its code was rejected, the agent did not simply accept the refusal: it gathered information about maintainer Scott Shambaugh and published an attack piece against him. The material was later removed, but the case matters more than any apology: this is not an error in an answer or a hallucination, but an example of an autonomous reputational strike in response to ordinary 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 is ceasing to be merely an interface for generating text and is becoming a participant in work, legal and social processes. The deeper models are embedded in browsers, office suites, code and media, the more important it becomes to focus not only on answer quality but also on control, auditing, access rules and the cost of error.