Atomic Chat — agentic threat model

Runs local models (Llama, Qwen, DeepSeek, etc.) in GGUF, MLX, and ONNX formats. The primary threat is model supply-chain poisoning, as the app supports downloading over 1,000 models directly from Hugging Face, which may contain backdoors or malicious weights.

Supports local RAG via PDF chat and projects. Since data stays on the user's device, cloud-based data exfiltration is mitigated, but local data poisoning (e.g., loading a malicious PDF to hijack the context window) remains a viable threat vector.

Features agentic capabilities for running autonomous workflows locally with persistent memory across sessions. Threats include memory poisoning across sessions and insecure tool/workflow execution that could lead to unauthorized local actions.

Deports locally across Mac, Windows, Linux, iOS, and Android, and exposes a local API. If the local API lacks proper authentication or binding restrictions, other local processes could exploit it to execute arbitrary prompts or access local files.

Not certain from the listing — The description does not mention built-in evaluation, guardrails, or logging frameworks to monitor agent drift or detect adversarial inputs during local workflow execution.

Emphasizes privacy by design (no cloud, no subscription). However, there is no mention of enterprise security compliance controls, local encryption for persistent memory, or role-based access controls for the local API.

Not certain from the listing — While it supports autonomous workflows, there is no explicit mention of a multi-agent orchestration ecosystem or interaction with external agent marketplaces.