Skip to main content

Security-focused AI workflow platform

TopFlow: govern and build secure AI

TopFlow is a visual platform for composing, validating, and running AI workflows. It brings model, data, and control-flow steps into a path teams can inspect and adapt. The OSV vulnerability scanner is one concrete use case—and a demonstration of keeping verified findings authoritative while treating model reasoning as an untrusted proposal.

Explore TopFlow Explore the Code

Inspect the proof

One implemented detection workflow, with source and tests you can review.

Source-backed vulnerability findings

The scanner parses supported repository manifests and queries OSV. Findings originate in source data, not model-generated advisory prose. Unsupported inputs and incomplete coverage remain limitations.

Read the scanner implementation

Untrusted Reasoning Worker (URW)

The model receives minimized finding fields and proposes structured selections. Code checks IDs against the known finding set before rebuilding the report from source records.

Read the URW implementation

Reviewable boundary tests

Tests exercise malformed selections, unknown finding IDs, and deterministic fallback. They are evidence for specific boundaries, not proof that every deployment is secure.

Inspect the URW regression tests

The factual path stays separate

  1. 1. Establish source findings. Parse supported manifests and obtain OSV records before asking a model to reason.
  2. 2. Constrain the proposal. Send minimized fields, not advisory prose, and request structured IDs and selections.
  3. 3. Validate and rebuild. Reject unknown IDs and reconstruct accepted findings from the source records, with deterministic fallback where needed.

A valid ID does not prove a good ranking. Optional model commentary is not authoritative evidence, and source coverage must be assessed independently.

Know which mode you are reviewing

Cached demo

Sample results explain the pattern without API keys. They are not a live scan of your repository or evidence of its current security posture.

Live detection

Real scans reach GitHub and OSV; authorized GitHub credentials may be required. Reports can use deterministic text without an AI key, or optional BYOK model reasoning.

Browser workflow editing does not mean server-free execution. Live requests reach server routes and configured providers; review credentials, logs, and hosting for your deployment.

Review TopFlow's license terms before reuse or deployment.

Explicit limits, not blanket assurances

  • Supported-manifest scanning does not detect every vulnerability or replace a security assessment.
  • Finding-ID validation does not establish ranking quality or the correctness of every generated comment.
  • General-purpose JavaScript sandboxing and human-gated outward actions are separate controls; do not infer them from a read-only scanner demo.
  • Workflow templates and exported code are reference material, not compliance attestations or production-deployment approvals.

Translate the pattern into an approval decision

Evaluate untrusted AI output

Five vendor questions, evidence to request, and the risks ID validation does not settle.

Make the control owner accountable

A CISO's view of defense-in-depth, operating evidence, and bounded pilot approval.

Discuss Your AI Security Architecture