Anthropic Is Loosening Claude’s Cyber Safeguards for Verified Defenders

Morning, {{first name | folks}}! AI is moving further into real-world systems this week, from giving security teams more freedom to use powerful models to Meta open-sourcing infrastructure software, smaller multimodal models, higher-end image generation, and AI assistants that can actually take actions for you.

Anthropic is expanding its Cyber Verification Program, giving vetted security teams access to Claude with fewer restrictions for defensive work, authorized red teaming and specialized testing. The highest tier covers systems such as power grids, flight systems, and financial infrastructure.

The program builds on Project Glasswing, where Anthropic says its models helped partners identify more than 129,000 verified software vulnerabilities between April and July, including over 33,000 rated critical or high severity. The company is now turning that work into a broader program for organizations that can prove they have the right security controls in place.

Meta has open-sourced Rebalancer, a system it has used for more than nine years to solve large-scale resource allocation problems across its infrastructure. It handles everything from placing servers and ML workloads to routing traffic across data centers, while balancing constraints such as capacity, fault tolerance and latency.

Rebalancer now solves roughly 40 million assignment problems every day across more than 30 different problem types. Meta is releasing the system under Apache 2.0 alongside Rebalancer Explorer, a debugging tool that helps engineers understand why the system made a particular assignment and test what happens when constraints change.

Google has released Gemini Nano Banana 2.1, a new image-generation and editing model that can work with up to 14 reference images in a single prompt and generate images at up to 4K resolution. It also improves prompt following, text rendering, and support for wide and panoramic images.

The model adds Google Search grounding and configurable thinking levels, giving developers more control over how it handles visual tasks. It’s available through the Gemini API, opening the door to more capable image generation for products, ads, design workflows, and other visual content.

Google DeepMind has released EmbeddingGemma 2, an open model that brings text, images, video, audio and code into a shared embedding space. Developers can now build cross-media search without maintaining separate models for each type of content.

The model is built to run locally and can use roughly 567MB of active RAM after quantization. That makes it an interesting option for private, on-device search, where the data stays on the device instead of being sent to the cloud.

Hark has released Hark Pro, a personal AI assistant designed to handle tasks across your digital life instead of just answering questions. It can work with email, calendars, files and connected services while using a computer-use model to navigate websites and complete tasks end-to-end.

One of its more interesting choices is making the agent’s browser activity visible while it works, rather than hiding what happens behind the scenes. Hark is also positioning privacy as a core part of the product, saying sensitive credentials are kept in an encrypted vault and that it does not sell user data or share it with advertisers.