PROMETHEUS solves the core security crisis of autonomous AI agents interacting with open-source repositories: unvetted package execution can compromise infrastructure in seconds. Built for the Vultr Hackathon Challenge 1 (Blast Radius Zero), PROMETHEUS enforces an architectural guarantee: discovery is not authority. Our system completely decouples repository inspection from execution. Untrusted code analysis is bounded by Vultr Serverless Inference (São Paulo region, sao) with strict token, cost ($0.000259/run), and rate limits. The host operates with zero public listening TCP ports, communicating exclusively over a local Unix domain socket with strict permissions (0660 root:prometheus). At Stage 03, an explicit Human Authorization Gate holds execution disarmed by default. When authorized, code runs inside an ephemeral, disposable sandbox with strict seccomp profiles, CPU and memory quotas, and zero host network access. When simulated hostile payloads attempt unauthorized network egress or file tampering, the broker immediately traps the violation, quarantines the container, and preserves host integrity. Every run generates cryptographically sealed SHA-256 evidence manifests matching our public GitHub repository. PROMETHEUS turns containment into a verifiable governance property, not a promise.
Category tags: