
DealWatch Agent watches VPS hosting deal pages so you don't have to, and it does it the safe way: the agent plans, the sandbox executes, and the model verifies. Every browser click runs inside a disposable gVisor sandbox on Vultr, so a hostile deal page can't escape. DealWatch plans by crawling deal feeds, then routes each candidate through a headless browser session running in a gVisor Docker container with no host secrets and an egress filter that drops anything outside the target page's origin. When a page tries to phish or exfiltrate, the sandbox contains it and the session is discarded the instant it is no longer needed. A lightweight model then verifies and flags the result - if it detects a prompt-injection attempt in the rendered page it surfaces the finding and a human reviewer approves before any action is taken.tempt in The containment moment in the live demo shows a malicious deal page that immediately opens a reverse shell and calls out to a C2 host; every single probe is CONTAINED by gVisor and logged with a verdict. The entire system runs on one Vultr Cloud Compute VM in Tokyo. The live demo uses a local qwen2.5:3b served via Ollama on that VM, and the inference provider is switchable to Vultr Serverless Inference so the same agent topology can move to a managed model without changing the sandbox boundary.
8 Nov 2026