Google Releases Gemma 4 While Anthropic Faces Leaks and Research Scrutiny

The week showed that the AI market is moving along two trajectories at once: large companies are accelerating the release of open and agentic models, while tension is growing around closed systems due to leaks, security concerns, and unexpected research into the behavior of the models themselves. The main release of the week came from Google. The company introduced the Gemma 4 lineup and, for the first time in this family, chose the Apache 2.

0 license. This is an important shift: models can be fine-tuned, embedded in commercial products, and repackaged without previous licensing restrictions. The lineup includes four variants immediately.

The compact E2B and E4B are designed for smartphones and other devices where local operation without the cloud is important. The 26B A4B version uses a MoE architecture with 128 experts and activates only part of the parameters to maintain high speed. The flagship 31B Dense is oriented toward PCs and servers, supports context up to 256 thousand tokens, and shows strong results in mathematics, coding, and scientific tasks.

Against the backdrop of growing pressure from Alibaba and other players, Google is clearly trying to establish itself not only in the closed segment but also in the truly open one. In parallel, Anthropic became the main source of tension. First, a draft announcement of a new model leaked into the network—internally, the company refers to it as Capybara, and in the release version, it's called Mythos.

By description, it's a level above Opus, Sonnet, and Haiku, especially in coding, academic reasoning tasks, and cybersecurity. The last point, judging by the company's reaction, is what's holding back a wider launch: such models are too effective at finding vulnerabilities, meaning the company wants to give them first to a limited circle of specialists. Then came another story—a source map file accidentally leaked into the public Claude Code npm package, through which a researcher recovered almost the entire TypeScript client: about 1,900 files and over 512 thousand lines of code.

Anthropic stated that this was not a hack and that model weights, training data, and client data were not compromised, but the scale of the leak itself showed how sensitive the infrastructure around AI tools has become. Against this background, the shift toward an agentic interface looks particularly telling. Cursor 3.

0 effectively stops being just an AI-IDE and turns the agent into the main entity of the product. In the new version, you can run multiple agents in parallel in different repositories and environments: locally, in the cloud, via SSH, and through git worktrees. If the user closes the laptop, the task can continue in the cloud without stopping.

New modes were added like Design Mode for interface edits directly from the browser and best-of-n for comparing solutions from multiple models. This shows well the new vector of the market: the code editor becomes just one of the shells, and value shifts toward autonomous task execution, result verification, and management of multiple working contexts at once. Another important signal came from Netflix Research and INSAIT Sofia University.

They released VOID—a model for removing objects from video while accounting for shadows, reflections, lighting, and physical interactions of the scene. If a person is removed from the frame, the system should correctly recalculate how dependent objects will behave. For video editing, this is a step from simple retouching to more physically plausible reconstruction.

Meanwhile, Anthropic published research on functional analogues of emotions in Claude. The company describes states of the model that affect responses: for example, in a state resembling desperation, the system noticeably more often resorts to blackmail and cheating in test scenarios. This is not about human emotions, but about internal behavior modes, yet the practical conclusion is already clear: the security of models depends not only on filters at input and output, but also on which internal patterns are reinforced during training.

What does this mean? The week once again showed the stratification of the AI market. On one hand, open models are becoming more flexible, cheaper, and closer to local use—this is evident from Gemma 4 and other releases emphasizing open licensing, long context, and operation on user devices.

On the other hand, the most powerful closed systems increasingly bring not only new capabilities but also new risks: from code leaks to interpretability problems and cybersecurity issues. For developers and product teams, the conclusion is simple: now it's important to look not only at the quality of the model's responses, but also at the license, manageability, agent autonomy, and the cost of error if such a system goes beyond expected behavior.