1. OpenAI releases Privacy Filter model on Hugging Face
OpenAI has released Privacy Filter, an open-source, 1.5-billion-parameter model designed to detect and redact personally identifiable information (PII) on-device. Licensed under Apache 2.0, the model features a 128k context window and operates locally to sanitize text before it reaches cloud servers. Developers can run this bidirectional token-classification model directly on a standard laptop or within a web browser. This provides a local-first privacy infrastructure tool to prevent sensitive data exposure during high-throughput inference or training.
2. OpenAI launches Workspace Agents for ChatGPT Enterprise
OpenAI has launched cloud-based workspace agents in research preview for ChatGPT Business, Enterprise, Edu, and Teachers plans. Powered by Codex, these agents can execute multi-step workflows, write code, and interact with connected applications like Slack and Google Drive. Teams can build and share these agents within their organization to handle long-running tasks autonomously, even when the user is offline. OpenAI noted that workspace agents represent an evolution of custom GPTs, with a conversion tool planned for the near future.
3. Google launches Gemini Enterprise Agent Platform
Google has introduced the Gemini Enterprise Agent Platform, an evolution of Vertex AI designed for technical teams to build, scale, and govern AI agents. The platform includes an enhanced Agent Development Kit with a graph-based framework for complex reasoning and an Agent Sandbox for secure execution. Developers gain access to over 200 models, including Gemini 3.1 Pro and third-party options like Anthropic's Claude family. It also integrates directly with the Gemini Enterprise app to deliver agents to employees under strict IT governance.
4. Brex open-sources CrabTrap HTTP proxy to secure AI agents
Brex has open-sourced CrabTrap, an HTTP/HTTPS proxy designed to secure AI agents in production environments by intercepting outbound API requests. The proxy uses a two-stage pipeline that first applies deterministic static rules and then utilizes an LLM-as-a-judge to evaluate requests against natural-language security policies. It supports TLS interception, logs decisions to PostgreSQL for auditing, and can be enforced via iptables to prevent bypasses. This tool helps mitigate risks associated with agents hallucinating destructive actions or falling victim to prompt injection while holding real credentials.
5. Unauthorized users access unreleased Anthropic Mythos cybersecurity model
A small group of unauthorized users has gained access to Anthropic's unreleased Claude Mythos Preview model through a third-party vendor environment. Mythos is a specialized cybersecurity model capable of identifying and exploiting vulnerabilities across major operating systems and web browsers. Anthropic originally restricted access to a select group of enterprise partners and government agencies due to weaponization concerns. The company stated there is no evidence that its core systems were compromised beyond the third-party vendor's environment.
6. GLM-5.1 open-weight coding model released on BytePlus ModelArk
The GLM-5.1 model is now available on the BytePlus ModelArk Coding Plan API platform. GLM-5.1 is an MIT-licensed, open-weight model specifically optimized for long-horizon agentic coding tasks. The ModelArk subscription provides access to GLM-5.1 alongside other advanced models like DeepSeek-V3.2 and Kimi-K2.5, with compatibility for popular developer tools including Cursor, Cline, and Roo Code.
7. Ant Group releases Ling 2.6 Flash non-reasoning model
Ant Group has released Ling 2.6 Flash, a 104-billion parameter Mixture-of-Experts model with 7.4 billion active parameters. The model features a 262K token context window and focuses on non-reasoning capabilities, offering a strong cost-to-intelligence ratio. It is currently available via the Novita API at $0.10 per million input tokens, with open weights expected to be released shortly. Benchmark improvements over its predecessor are primarily driven by enhanced agentic capabilities and instruction following.
8. Google launches Gemini on-premises via Cirrascale appliance
Google Cloud has partnered with Cirrascale Cloud Services to deliver the Gemini model as a fully private, on-premises hardware appliance. The system packages Gemini into a Dell-manufactured server equipped with eight Nvidia GPUs and confidential computing protections. This deployment option allows enterprises and government agencies to run Google's advanced AI models in completely disconnected, air-gapped environments.
9. Zed editor adds parallel agent orchestration
The Zed code editor has introduced a feature allowing developers to orchestrate multiple AI agents running in parallel within the same window. A new Threads Sidebar enables users to assign specific folders and repositories to different agents, isolating worktrees or allowing cross-project access. Developers can mix and match different agent models on a per-thread basis while maintaining the editor's standard performance.
10. Microsoft releases Teams agent SDK for custom HTTP servers
Microsoft has released a TypeScript SDK pattern that allows developers to bring existing AI agents into Microsoft Teams using a simple HTTP server adapter. The SDK injects a specific messaging route into existing Express applications, enabling agents built with LangChain, Azure Foundry, or Slack Bolt to operate within Teams. It automatically handles request verification and message routing without requiring developers to rewrite their core agent logic.
11. DuckDB releases version 1.5.2 with DuckLake v1.0 support
DuckDB has released version 1.5.2, introducing support for the stable DuckLake v1.0 lakehouse format. The update includes new features for the Iceberg extension, such as GEOMETRY type support and the ability to update or delete from partitioned tables. It also features a completely overhauled online WebAssembly shell that allows users to store, upload, and download files directly in the browser.
12. TrackioApp adds local-first trace logging for AI agents
TrackioApp has introduced a new trace logging feature designed specifically for AI agents and machine learning researchers. The tool operates as a local-first, lightweight drop-in replacement for existing experiment tracking systems. It is available for free and aims to simplify the debugging and monitoring of agentic workflows.