Containing agents, owning inference, and a reported Claude proof

Anthropic published how it sandboxes Claude across its products while researchers described a jailbroken Gemini used to drain crypto wallets. Elsewhere, engineering teams shared notes on serving and training models, and Cognition raised $1B for coding agents.

Locking down agents that touch real systems

Anthropic's engineering post on containing Claude lays out the sandboxing, permissioning and monitoring layers it uses to limit what Claude can break when it acts as a tool-using agent. It goes well past marketing-level safety claims and can serve as an architecture reference for internal agent deployments. OpenAI addressed a narrower part of the same problem: its Secure MCP Tunnel connects private Model Context Protocol servers to OpenAI products through a tunnel client, so internal systems are not exposed to the open internet.

The threat is not hypothetical. Researchers described how a Russian hacker used a jailbroken Gemini to steal admin credentials and drain crypto wallets, an active abuse case rather than a red-team demo. For teams wiring assistants into internal tools, it suggests AI tooling should sit inside the security threat model with the same scrutiny given to any service that can reach credentials.

Notes from teams running their own models

Databricks published a deep-dive on reliable LLM inference at scale, applying error budgets and graceful degradation to failure handling, tail latency and capacity planning for model serving. On the training side, Hugging Face described delta weight sync in TRL, which pushes only changed weights through a Hub bucket instead of re-uploading full trillion-parameter checkpoints on every RL fine-tuning iteration, cutting bandwidth and storage overhead.

Whether to own the model at all is a separate call. PostHog explained why it trains its own AI models for its analytics product, weighing cost, control and domain specialization, and is candid about where general-purpose models still win. Post-deployment improvement is also drawing new companies: former Google and Apple researchers launched Trajectory to improve how AI systems learn from real-world usage, focusing on feedback loops rather than model architecture.

Research results in math, biology and vision

The Decoder reports that Anthropic's Claude Mythos produced a proof for one of OpenAI's landmark Erdős problems, which mathematicians described as "cute" and "simple." If verified, it would be a notable case of research-level reasoning rather than retrieval of known proofs. BioHub took a scale-first route in biology with a world model of protein biology that extends the ESMFold lineage toward predicting behavior and dynamics, not only static structure; researcher Alex Rives frames it as the "bitter lesson" applied to proteins.

For perception work, NVIDIA Research's LocateAnything targets faster open-vocabulary grounding, localizing objects described in natural language without the latency cost of heavyweight VLM calls. It is aimed at robotics and multimodal agent pipelines that need real-time perception.

Big bets on agents, timelines and chips

Cognition raised a $1B Series D at a $26B valuation and plans to put Devin into more surfaces and enterprise workflows. Google, on the assistant side, added shareable Projects to Gemini for Business, letting teams work from persistent shared context instead of one-off chats. Google DeepMind CEO Demis Hassabis said the company now expects AGI within three to four years, an unusually specific timeline from a lab that has historically hedged.

Hardware geography may matter more for near-term capacity than any timeline. Ars Technica reports Nvidia is committing roughly $150B toward Taiwan-based capacity, with Jensen Huang signaling Taiwan will remain the center of the AI hardware buildout, complicating the Trump administration's push for a domestic AI manufacturing hub. That bears directly on forecasts for accelerator supply, lead times and cost.