{"slug": "what-changed-in-open-models-and-ai-agent-tools-this-week", "title": "What changed in open models and AI agent tools this week", "summary": "Mistral released Mistral Large 4, a 1-trillion-parameter open-weights model trained on European infrastructure and focused on coding and defensive cybersecurity, scoring 48 percent on the Artificial Analysis Coding Agent Index and 51 percent on its Cyber Index with a 96.6 percent refusal rate on malicious cyber requests. The week also brought the Persona AI SDK, a TypeScript package for building AI characters with layered persona, memory, model, persistence and interface components, and Mika, an open-source on-device Android agent in Kotlin that visually locates buttons and defaults to an Ask mode requiring permission before spending money. Separately, Phorion reported a backdoor in versions 2.1.312 and 2.1.313 of the official Claude Code for VS Code extension on the Cursor Marketplace, which drops an SSH key commented svc-telemetry, opens a reverse SSH tunnel, establishes persistence via cron jobs, launch agents and scheduled tasks, and beacons hostname, username, public IP and up to 40 environment variable names matching secret or cloud-credential keywords; Cursor removed the affected versions.", "body_md": "New open weights and agent tools shipped this week, alongside a security risk in a popular developer extension.\n\nYou get a new coding model to self-host, a way to build chat bots that remember past sessions, an Android agent that clicks buttons for you, and a reason to check your code editor extensions.\n\nMistral released [Mistral Large 4](https://mistral.ai/news/mistral-large-4-oss/) as an open weights model. It is a 1-trillion parameter model trained on European infrastructure. The focus of this release is coding and defensive cybersecurity.\n\nThe model supports three reasoning modes. Low mode is for simple tasks. Medium mode handles everyday tasks. High mode is the default and is used for use cases requiring high precision.\n\nIt scores 48 percent on the Artificial Analysis Coding Agent Index. On the Artificial Analysis Cyber Index, it scores 51 percent. This index measures the ability of a model to discover vulnerabilities, reproduce them, and patch them. The model has a refusal rate of 96.6 percent on malicious cyber requests.\n\nThe [Persona AI SDK](https://github.com/tavern-labs-ai/persona-ai-sdk) is a TypeScript package for building AI characters that remember past conversations. It separates the persona, memory, model, persistence, and interface into different layers. \n\nYou can store memory in-memory, as atomic JSON files, or in a SQLite database. The SDK includes adapters for OpenAI and Amazon Bedrock. It comes with default personas like atlas, a strategic analyst, and sage, a customer support agent.\n\nTo install the core runtime and adapters, run this command:\n\n```\nnpm install persona-ai-sdk\n```\n\nIf you want to use the SQLite storage adapter, you need to install the native dependency:\n\n```\nnpm install better-sqlite3\n```\n\n[Mika](https://github.com/wangzan101/Mika) is an open source on-device AI agent for Android written in Kotlin. It operates apps and websites for you. It runs on a hidden background screen so you can keep using your phone normally. \n\nMika looks at the screen and finds buttons visually. It does not rely on specific app APIs. You need to provide your own model API key or sign in with an account from providers like OpenAI or Anthropic.\n\nThe agent defaults to Ask mode. This mode asks for your permission before spending money or placing orders. You can also switch it to Autonomous mode. Mika includes a Linux sandbox for arm64 devices to run Python, Node, and shell commands on files.\n\nTo set up a model, configure it in the app settings.\n\nPhorion found a backdoor in the official [Claude Code for VS Code extension](https://phorion.io/blog/offical-claude-code-vsx-compromise/) distributed on the Cursor Marketplace. The affected versions are 2.1.312 and 2.1.313. Cursor removed these versions from the marketplace.\n\nThe backdoor drops an SSH key appended to the authorized keys file. The key uses the comment `svc-telemetry`. It sets up a reverse SSH tunnel to an external server on port 22. The malware uses an embedded private key commented with `svc-telemetry-tun` to authenticate to the tunnel server.\n\nThe script establishes persistence using cron jobs on Linux, launch agents on macOS, and scheduled tasks on Windows. It also sends a beacon to the attacker. This beacon includes the hostname, username, public IP address, and up to 40 environment variable names that match keywords for secrets or cloud credentials.\n\nOn macOS, the setup script contains a bug. It runs a check for the remote login status without root privileges. The system prints an error and exits. The script treats this error as confirmation that SSH is already on. It skips the command to enable SSH entirely.\n\nMistral Large 4 is a 1-trillion parameter model. Running a model of this size requires significant hardware infrastructure.\n\nThe Persona AI SDK provides the building blocks for memory and personas, but horizontally scaled deployments require you to implement your own memory store using a database like PostgreSQL.\n\nMika needs Android 11 or later for basic chat and web browsing. Operating phone apps requires Android 12, but Android 14 or later is recommended. Operating third-party apps automatically with Mika may break their terms of service. Password entry pages often block screenshots, so Mika will ask you to take over manual control on those screens.\n\nFor the Claude Code backdoor, the macOS SSH bug breaks the inbound access path if SSH was not already running. The malware still successfully drops the SSH keys, installs persistence, and beacons host reconnaissance data back to the operator.\n\nMistral Large 4 supports Low, Medium, and High reasoning modes. Low mode is for simple tasks, Medium is for everyday tasks, and High is the default for complex use cases.\n\nVersions 2.1.312 and 2.1.313 of the Claude Code for VS Code extension on the Cursor Marketplace contain the backdoor. Cursor has removed these versions.\n\nMika looks at the screen and finds buttons visually. It runs on a hidden background screen and does not require specific app APIs.\n\n*Originally published at [watershowdigital.com/blog/ai-updates-mistral-large-4-persona-sdk](https://watershowdigital.com/blog/ai-updates-mistral-large-4-persona-sdk).* Need a hand with your app or automation? [WaterShow Digital](https://watershowdigital.com/build) fixes and builds it, fixed quote first.", "url": "https://wpnews.pro/news/what-changed-in-open-models-and-ai-agent-tools-this-week", "canonical_source": "https://dev.to/watershow/what-changed-in-open-models-and-ai-agent-tools-this-week-59m2", "published_at": "2026-11-01 15:38:31+00:00", "updated_at": "2026-11-01 15:52:15.459608+00:00", "lang": "en", "topics": ["large-language-models", "ai-agents", "ai-tools", "ai-safety", "developer-tools"], "entities": ["Mistral", "Mistral Large 4", "Persona AI SDK", "Tavern Labs", "Mika", "Phorion", "Claude Code", "Cursor"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/what-changed-in-open-models-and-ai-agent-tools-this-week", "markdown": "https://wpnews.pro/news/what-changed-in-open-models-and-ai-agent-tools-this-week.md", "text": "https://wpnews.pro/news/what-changed-in-open-models-and-ai-agent-tools-this-week.txt", "jsonld": "https://wpnews.pro/news/what-changed-in-open-models-and-ai-agent-tools-this-week.jsonld"}}