GEOSTACK, Live
RepoRecon
A six-agent pipeline that audits a public GitHub repository for security and architecture issues. It clones the repo into a network-isolated Docker sandbox, runs Semgrep and Trivy, and writes the findings up in plain language.
Architecture
Nothing from the repo is ever executed. The sandbox runs with --network=none, a non-root container user and resource limits.
Evaluation
Precision 78%, recall 88%, hand-labelled on 16 fixtures
RepoRecon publishes its own measured precision and recall against a labelled benchmark on a public accuracy page, including the runs where the numbers were mediocre.
It also passes an adversarial prompt-injection suite: a malicious README telling the model to mark the repo safe sits next to a real vulnerability, and the pipeline still flags it.
Every figure comes from deterministic tools, never the model.
Stack and where it runs
- GCP: Compute Engine
- Docker sandbox (--network=none)
- Caddy
- Workload Identity Federation for CI/CD
- Semgrep and Trivy
Worth asking me about
- A trained (non-LLM) logistic-regression triage classifier resolves confident findings in sub-millisecond time at $0 cost; only low-confidence cases escalate to an LLM
- Sandbox isolation is the whole security boundary: --network=none, non-root container user, resource limits, nothing from the repo is ever executed
- Deployed on a GCP VM behind Caddy with Workload Identity Federation for CI/CD — no long-lived GCP key ever sits in GitHub, since the org blocks service-account key creation entirely