NativeMind: Your fully private, open-source, on-device AI assistant

The extension is designed to provide a private, on-device AI assistant in the browser by connecting to local LLMs (Large Language Models) via Ollama. Its intention is to run AI models locally, without sending data to the cloud, and to support a wide variety of open-source models.

Key behaviors and APIs:

• Model Management: Maintains a large list of supported models, each with metadata (URLs, VRAM requirements, etc.), and fetches models from trusted sources (e.g., HuggingFace, GitHub).
• Local Storage: Uses browser storage APIs (IndexedDB, Cache API) to cache models and artifacts for offline/efficient use. No evidence of exfiltration or suspicious storage usage.
• WebGPU Usage: Leverages WebGPU for on-device inference, including shader compilation and GPU memory management. No evidence of GPU abuse or crypto-mining.
• Network Activity: Downloads models and WASM binaries from open, reputable sources. No evidence of arbitrary or dynamic remote code fetching or C2 communication.
• Filesystem Activity: Uses browser sandboxed storage only; no access to local disk outside browser APIs.
• Process Execution: No evidence of launching external processes or using native messaging.
• RPC/IPC: Uses structured messaging for internal communication (e.g., birpc), not for external code execution.
• Logging: Uses a logging library for debugging and warnings, not for data exfiltration.

No strong indicators of malicious activity were observed.

Obfuscation:

• The code is minified (short variable names, no whitespace), but not obfuscated. There is no evidence of string encryption, control flow flattening, or other advanced obfuscation techniques.

Summary:

• The extension's code is consistent with its stated purpose: providing private, on-device AI via local models. All network activity is transparent and directed at model repositories. No suspicious or malicious patterns were detected.