LLM 0.37.2 fixes a security vulnerability reported by Daniel de Araujo. The issue arises from the extra-openai-models.yaml mechanism, which is now managed by the new "llm openai endpoint add" command. This mechanism allows using LLM with OpenAI-compatible API endpoints from other providers. However, Daniel's report highlighted two scenarios where the configured OpenAI API key could be accidentally sent to one of those providers: if the key had never been stored, or if a responses:true model was run with -o chat_completions 1. The update addresses these concerns, enhancing security for developers and operators of AI agents.

Pomerado is a self-serve platform that enables developers to create APIs for any website in minutes, allowing agents to interact with legacy systems. The platform uses a natural language description and a few questions about the workflow to generate a Machine-Callable Protocol (MCP) in under an hour. The MCP is a deterministic, self-healing API that returns results in 5-30 seconds. Pomerado handles authentication, 2FA, and site changes, and includes a maintenance job to repair and improve the tool if future runs fail. This technology matters to people building or operating AI agents because it overcomes the limitation of agents being unable to interact with legacy systems, which is a common bottleneck in AI deployments, particularly in verticals like healthcare, professional services, and telecom.

Gemini 4 Carbon is a new version of the Gemini AI model, a conversational AI developed by DeepMind. The update introduces several key changes, including improved performance on long-form dialogue, enhanced knowledge retrieval, and better handling of multi-turn conversations. Gemini 4 Carbon is also more energy-efficient than its predecessor, Gemini 4, and can be deployed on a wider range of hardware configurations. The update is significant for developers and operators of conversational AI systems, as it enables more accurate and efficient interactions with users. The improved performance and energy efficiency make Gemini 4 Carbon a valuable tool for building and deploying conversational interfaces in various applications, such as customer service chatbots, virtual assistants, and language translation systems.

Pots is an open-source AI coworker that integrates with Slack, Discord, Telegram, and other chat platforms, allowing teams to use it as a middle layer to streamline their workflow. Unlike Grok Bot, Pots is designed to work within existing channels and workflows, providing features tailored to each company's needs. It can listen to team discussions, build prototypes, and deploy pages without requiring setup. Pots is built on a more open architecture, allowing users to choose any model and deploy it on their own servers, giving them greater flexibility and control over their data.

Mimzie is a free desktop app that provides a GUI for managing knowledge work agents from multiple vendors, including ChatGPT, Claude, Gemini, and Grok. It allows users to run these agents in their native CLI harnesses, eliminating the need for agent-specific interfaces. Mimzie is designed to be agent-neutral, meaning users can utilize any AI plan or combination of plans without incurring additional subscription fees. Key features include local file editing with cloud syncing, reviewable changes, and integration with bots from various AI agents.

Gemini Decision 1 is now available on AI Gateway, a unified platform for accessing and managing AI models. This multimodal decision model can process text, images, audio, video, and PDF documents, providing Boolean probabilities, choices, and scores. It can be used for tasks such as routing requests, checking policies, filtering agent tools, and ranking results without generating output tokens. The model is accessible through the AI SDK, TypeSafe-compatible API, or Decisions API, with varying input support depending on the API used. AI Gateway tracks usage and cost, including decision calls, and provides a centralized way to manage models.

Google has released Gemini 0.36, a new version of its large language model (LLM) with improved capabilities. The update is available to the general public, not just a subset of cybersecurity professionals as previously announced. The new model is priced at $2 per million input tokens and $10 per million output tokens at launch, but will double in price to $4 and $20 respectively on December 1st. Gemini 0.36 offers three thinking effort levels, with a default medium setting. The model can be used for tasks such as generating HTML code for animations and creating open-world 3D games using Google's Antigravity IDE. This release is notable for its improved capabilities and the ability to be used by a broader range of users.

Gemini 4 Carbon, a large language model developed by Google, is now available on the AI Gateway platform. This model is designed for complex tasks such as software engineering, cybersecurity defense, and sustained agentic work. Key features include a 1M-token context window, improved coding and scientific reasoning abilities, and enhanced cybersecurity capabilities. AI Gateway provides a unified API for integrating and managing models like Gemini 4 Carbon, offering features such as custom reporting, budgeting, and routing rules. The model is currently available at 50% off until November 30, 2026, and can be accessed through the Vercel CLI and various coding agents.

Anthropic released version 0.31 of their llm-anthropic, which includes Fable 5.2 and Mythos 5.2 models. The update allows users to run prompts through Fable 5.2 using the llm command, enabling generation of visual media such as images. Testing revealed that the Fable 5.2 model, particularly at the xhigh and max reasoning levels, sometimes runs out of tokens when generating complex images, resulting in failure. This issue is similar to a previously observed problem with the Opus 5.5 model but is new to the Fable series.

ApiClay is a platform that enables users to discover, create, and utilize APIs from various sources, including websites, Google Sheets, databases, and HTTP services. Key features include: * API catalog search and playground for inspection and testing * Ability to create APIs from non-API sources using Website to API * Support for importing existing OpenAPI/Swagger files * Customizable API endpoints with control over access, trials, plans, usage limits, and pricing * Aggregate multiple providers into a single response for comparison-style products * Exposed MCP endpoint for AI agents like Claude and ChatGPT to search and call APIs * Practical data tools for web scraping, enrichment, monitoring, cloud browsing, and extraction * Encrypted credentials for publishers and paid calls charged per request * Free browsing and reading of API docs, with paid calls visible before use

Thrum is a tool that enables developers to run multiple AI coding agents, such as Claude Code, Codex, and Copilot, as a single team on their local machine. Key features include runtime agnosticism, allowing users to mix and match different agents, a unified coordinator for all projects, and extended agent memory and state management. Thrum also provides a searchable conversation history and supports semantic search across conversations. This tool is relevant to developers using AI agents in their workflows, as it offers a centralized and scalable solution for managing multiple agents and their interactions.

Vercel Container Registry has increased the maximum size of compressed image layers from 500 MB to 2 GB, while maintaining a total image size limit of 15 GB. This change allows for more efficient storage and management of container images, particularly for large applications. The new support for larger compressed image layers is available in Vercel CLI v58.3.0 and later, and enables teams to manage repositories, images, and tags with the `vercel vcr` commands.

Token Pass now includes GLM 5.3 Flash, a native multimodal coding model with a 1M-token context window, in addition to existing models Kimi K3, GLM 5.3, and DeepSeek V4.1 Flash. This update is available to Pro teams with a $30/month subscription, which provides $80 in monthly credit and is accessible through AI Gateway. The new model can be added from the Token Pass page in the AI Gateway dashboard or by referring to the Token Pass documentation for setup. This change matters for developers building or operating AI agents, as GLM 5.3 Flash offers an additional coding model option for integration with their projects.

The Workflow SDK on Vercel now uses WebSockets by default for event and stream writes, replacing individual HTTP requests. This change is effective with the upgrade to [email protected] or later and does not require code modifications. As a result, workflows with many steps will make fewer HTTP requests, improving performance and reducing costs. A benchmark showed that the new WebSocket approach delivered a full response 4.2 times faster than the previous implementation, with no chunk delayed over a second.

Meta's Muse Ember reasoning model is now available on AI Gateway, a unified API platform for calling models. Muse Ember is designed for high-volume tasks, such as coding sub-tasks, tool calls, and document analysis, and can handle text, image, and video input within a 1M-token context window. The model can be called through various APIs, including the AI SDK, Chat Completions, Responses, and Anthropic Messages APIs. AI Gateway provides a unified interface for tracking usage and cost, configuring retries and performance optimizations, and includes built-in custom reporting and Zero Data Retention support.

Anthropic's Claude Fable 5.2 is now available on AI Gateway, a unified API for calling models. Fable 5.2 improves upon previous models in coding, reasoning, and vision, enabling long, unattended tasks and acting as an orchestrator or advisor. It also communicates more clearly and generates output faster. Fable 5.2 requires data retention, but Anthropic does not support Zero Data Retention (ZDR) for this model. When ZDR is enabled, AI Gateway returns an error instead of serving the request. To work around this, use model fallbacks by adding a models array to providerOptions.gateway. This allows AI Gateway to try other models when the cybersecurity classifier declines a request.

Three new AI security audit skills have been released for Agent Skills, allowing for static analysis of AI models without executing them. The skills, electron-surface-audit, llm-app-audit, and invisible-char-audit, analyze desktop applications, source code, and text files to identify potential security vulnerabilities. Unlike traditional security tools, these skills do not execute the target, reducing the risk of compromising the system. Each skill provides detailed findings, including file and line numbers, allowing developers to verify and address potential issues. The skills are compatible with Claude Code, Claude Desktop, and other assistants that support SKILL.md, and can be installed via npm or cloned from GitHub.

The amla-sandbox Python package has been released, providing a secure, sandboxed environment for AI agent code execution. Key features include capability enforcement, a virtual filesystem, and network restriction, preventing shell escapes. The package allows developers to define constraints on tool usage, such as specific methods, parameters, and call limits. It is installable via pip, without requiring Docker or a virtual machine, and supports multiple languages including JavaScript and shell. The package includes a simple API for creating sandboxed environments and tools, as well as a constraint DSL for defining access controls.

Pentest-AI is an open-source AI-powered penetration testing and bug-bounty tool that separates proven findings from guesses by connecting an AI client or model to a real pentest workflow. The tool uses machine oracles to verify vulnerabilities, labeling findings as candidates or verified based on the oracle's decision. The AI model drives the investigation, but the oracle determines whether a finding is verified, preventing the model's guesses from being mistaken for results. Pentest-AI is designed for authorized bug bounty programs, allowing researchers to test systems they own or have explicit permission to assess, with a built-in dry run feature and report drafting from verified findings only.

Clawpatrol is a security firewall for AI agents that sits between the agent and its production environment, parsing traffic at the wire and enforcing rules written in HashiCorp Configuration Language (HCL). It can block malicious actions, such as destructive SQL queries or kubectl delete requests, and requires human approval for certain operations. Clawpatrol supports conditions based on protocol-specific facts, such as SQL verbs and table names, Kubernetes resources, and HTTP method and path. It offers three deployment options and can be used to secure agent traffic for various protocols, including Postgres, ClickHouse, and Kubernetes.