Skip to content

PROJECT.md

Latest commit

 

History

History
110 lines (93 loc) · 2.85 KB

PROJECT.md

File metadata and controls

110 lines (93 loc) · 2.85 KB
Raw

Provn: CLI-First AI Security Scanner

One-liner

Provn is a local pre-commit security scanner that blocks secrets, system prompts, and proprietary IP before they leave your machine.

Current Scope

Provn is CLI-only for now.

Included now:

  • Rust CLI for staged diff scanning
  • Layer 1 regex and entropy detection
  • Layer 2 AST-based checks
  • optional local Layer 3 semantic classification
  • HMAC-chained audit log
  • packaging, install flow, and CI support

Postponed:

  • bridge services
  • 3D office / Claw3D
  • monitoring dashboards

Why This Direction

  • It keeps the product focused on the real wedge: blocking unsafe commits
  • It reduces demo risk and maintenance overhead
  • It lets us lock speed, accuracy, and auditability before adding visualization

Architecture

git commit
   |
   v
Provn CLI (Rust)
   |
   +-- Layer 1: regex + entropy
   +-- Layer 2: AST / taint-style checks
   +-- Layer 3: optional local Gemma review
   |
   +-- verdict: allow | warn | block
   +-- auto-redact preview
   +-- HMAC audit chain

Tech Stack

  • CLI Core: Rust
  • Parsing: tree-sitter
  • Audit: HMAC-SHA256 + Ed25519
  • Semantic Layer: local Gemma 4 E2B
  • Fine-tuning: Modal + Unsloth + LoRA
  • Packaging: npm wrapper + Homebrew tap

What We Are Not Building Right Now

  • 3D office
  • websocket bridge
  • fleet management UI
  • cloud scanning by default
  • compliance dashboards

Target Metrics

  • Clean commit latency: p50 < 120ms, p95 < 200ms
  • Recall on seeded leaks: >95%
  • False positive rate: <2%
  • Bypass and audit events: logged and verifiable
  • Layer 3 must remain optional and fail-safe

Execution Plan

Phase 1: Core CLI

  • keep staged diff scanning fast and deterministic
  • improve regex coverage and entropy heuristics
  • harden AST checks for Python / TypeScript / JavaScript
  • keep the hook standalone and reliable

Phase 2: Audit + Semantic Layer

  • maintain tamper-evident audit records
  • keep local model serving optional
  • only escalate ambiguous findings into Layer 3
  • skip Layer 3 cleanly if it is slow or unavailable

Phase 3: Packaging + Proof

  • tighten LeakBench coverage and reporting
  • keep CI focused on the CLI
  • improve install and upgrade paths
  • prepare a terminal-first demo and README

Key Files

Provn/
├── provn-cli/
│   ├── Cargo.toml
│   ├── src/
│   │   ├── main.rs
│   │   ├── audit.rs
│   │   ├── config.rs
│   │   ├── diff.rs
│   │   ├── policy.rs
│   │   ├── redact.rs
│   │   └── scanner/
│   └── tests/
├── aegis-model/
│   ├── modal_finetune.py
│   ├── leakbench_train.jsonl
│   └── leakbench_eval.jsonl
├── scripts/
└── homebrew-tap/

Current Rule

If work does not directly improve the CLI, audit chain, local semantic layer, packaging, or CLI demo reliability, it is out of scope for now.