Orgo — agentic threat model

Not certain from the listing — Orgo provides the desktop infrastructure rather than the foundation models themselves, though it integrates with models like Claude. A key L1 threat is visual prompt injection, where malicious content on a web page or screen viewed by the model manipulates its desktop actions.

Not certain from the listing — Orgo focuses on VM execution environments rather than structured RAG or vector databases. However, data exfiltration is a major threat as agents can freely download, upload, or copy sensitive data within the desktop environment.

Orgo enables direct OS-level tool execution (mouse clicks, typing, terminal commands). The primary threat is tool misuse, where the agent executes destructive bash commands, fills out unauthorized forms, or navigates to malicious sites due to planning failures or adversarial prompts.

This is Orgo's core layer. It provides isolated Ubuntu VMs booting in under a second. Key threats include VM escape vulnerabilities, lateral movement to other customer VMs, and resource exhaustion (DoS) on the hosting infrastructure by rogue agent loops.

Not certain from the listing — There is no mention of built-in session recording, audit logs, or real-time guardrails to monitor and intercept harmful agent actions on the virtual desktop.

Not certain from the listing — While Orgo claims 'secure VMs', it does not specify compliance certifications (e.g., SOC2), identity and access management (IAM) controls for VM access, or policy enforcement mechanisms.

Not certain from the listing — The platform is designed for single-agent desktop automation and does not explicitly detail multi-agent orchestration, agent-to-agent trust boundaries, or marketplace risks.