Basalt: an open-source purple-team of AI agents that proves vulnerabilities by execution

AI security scanners have a trust problem

Most "AI security scanners" hand you a list of maybe-bugs a model thinks it found. You still have to check every one by hand. That is noise, not security.

Basalt proves findings by execution, not opinion

Basalt is an open-source tool where a purple-team of AI agents audits code, and a finding is only marked confirmed when its proof-of-concept is reproduced by execution in a clean, network-less sandbox — or matched against an authoritative vulnerability database. Every finding carries honest provenance: tool-proven, OSV-match, pattern-match or model-asserted. And every report shows the full verification funnel — from N candidates down to the few that actually reproduce — so you see the filtering, not just the conclusions.

How it works

The pipeline runs in stages: recon maps the attack surface; several independent red agents look for vulnerabilities and write a runnable exploit, with consensus voting to filter hallucinations; a skeptical purple judge demands a credible proof; execution reproduces that proof in an isolated Docker sandbox — a finding is confirmed only if the exploit actually fires; and finally a blue agent proposes a minimal patch that a grader validates by execution (green build, closed exploit, no regressions).

What is solid today

The deterministic scanners are stable and useful on their own: dependency N-day detection against OSV.dev, Go reachability via govulncheck, and cross-language hardcoded-secret detection (it never stores the secret value). It covers Python, Go, TypeScript/JavaScript, Rust and Solidity, and is built for CI with severity-gated exit codes, changed-files-only audits and cost estimation. The LLM discovery path — finding logic bugs with the red agent — is still beta: verified by execution, but coverage and precision are maturing.

Open, honest — and screening, not a replacement for a manual audit

Basalt is MIT-licensed and available on PyPI as basalt-scan. It is designed as automated screening, not a substitute for a professional manual audit of code that holds real funds — exactly the kind of deep review we do at Obsidiaan for smart contracts and critical systems. Building Basalt is part of how we push our own security and AI research forward.

Explore it on github.com/Arkessiah/Basalt.