ProofBox

**ProofBox** is a secure AI agent execution workbench built for Vultr's Challenge 01: Blast Radius Zero. It addresses two critical risks in autonomous AI systems: unverified execution claims and host exposure caused by running LLM-generated code or browser sessions without proper isolation. Deployed on **Vultr Cloud Compute (`vc2-4c-8gb`, Singapore)** and powered by **Vultr Serverless Inference**, ProofBox provides three execution patterns: - **Python Sandbox:** Generates and executes Python 3.12 data analysis scripts inside unprivileged, read-only Docker containers with network access disabled. It supports automatic error recovery and SVG chart generation. - **Browser Security Inspector:** Uses SSRF protection to block cloud metadata, localhost, private network addresses, and local file access before launching isolated browser inspections. It also audits DOM elements, links, and HTTP security headers. - **Red-Team Containment Matrix:** Allows users to test real adversarial scenarios, including secret exfiltration, sensitive file access, metadata SSRF, and memory exhaustion. A two-layer defense combines `nemotron-3.5-content-safety` with Linux `cgroupv2`, network namespaces, and DAC restrictions. Every execution generates a **ProofReceipt** containing SHA-256 hashes of the input, executed code, stdout, and output artifacts, alongside an HMAC-SHA256 signature and verified container teardown status. Our three-model inference pipeline uses `nemotron-3.5-content-safety` for pre-execution screening, `deepseek-v4.1-flash-1017` and `glm-5.3-flash` for code generation and self-healing, and `qwen3.8-27b` for output verification. **Results:** 12/12 adversarial tests passed (100%), with 3.29s average end-to-end latency, 41ms SSRF validation, zero host secrets injected, and verified container cleanup across test runs. ProofBox demonstrates how AI-powered automation can operate within strict execution boundaries while producing verifiable evidence of its actions.