Security & Data Handling Specification
Zero-Trust Execution by Architecture, Not Policy Promises
Granting read access to proprietary source code requires hard architectural guarantees. Nulltrap is designed so customer source code never persists on disk, never crosses tenant boundaries, and never trains foundation models.
Ephemeral Single-Tenant MicroVMs (`tmpfs`)
Every Deep Hunt and Delta Guard run provisions a dedicated, hardware-virtualized microVM. Target repositories are cloned exclusively into volatile memory (`tmpfs`). Upon run completion—or timeout—the microVM is terminated and memory pages are zeroed. No repository snapshots, build caches, or embeddings are persisted after report generation.
Two-Phase Network Air-Gap
Execution is strictly split into two network states:
- • Phase A (Dependency Fetch): Allowlisted read-only egress to standard package registries (`npmjs`, `pypi`, `crates.io`, container registries) to compile harnesses.
- • Phase B (Hunt & Exploit Verification): External network namespace egress is severed completely. Synthesized PoCs execute solely against loopback services inside the microVM—making accidental contact with live production APIs or external data exfiltration physically impossible.
Zero Data Retention (ZDR) & Zero Model Training
Private customer engagements run exclusively over enterprise Zero-Data-Retention inference APIs. Neither Nulltrap nor upstream model providers log, retain, or train models on your source code, prompts, or vulnerability findings. Bring-Your-Own-Key (BYOK) via AWS Bedrock or Google Vertex AI is supported for custom enterprise scopes.
Least-Privilege GitHub Scope & Mutual NDA
Nulltrap requires read-only repository contents access on scoped repositories (plus optional Security Advisory draft / Pull Request permissions if you want automated patch PRs opened directly). Every engagement includes a standard counter-signed Mutual NDA prior to repository connection.