BadgeIn x Skybridge

The description states that BadgeIn x Skybridge is a system composed of two components:

• BadgeIn: A human approval interface that shows the exact app, action, resource, destination, payload digest, expiry, and invocation count. It allows denial or one-time approval.
• Skybridge: A runtime enforcement layer that checks for valid, single-use grants before allowing actions to proceed.

The system uses:

• Python-based authority service
• Static HTML/CSS/JavaScript interface
• HMAC-authenticated exact grants with atomic one-use consumption
• Session-sandbox executor
• Cloudflare Tunnel for HTTPS demo access
• Qwen LLMs as fallbacks in a redundancy ladder

Inference: The product is described as a security-focused tool to prevent unauthorized AI actions by enforcing human authority at runtime.

The description states:

• AI agents are moving into business systems (calendars, files, inboxes).
• The weak point is not model accuracy but authority.
• Prompt injection succeeds when language can impersonate permission.
• The system assumes a model may be fooled and prevents that mistake from becoming an action.

Claim: "TFB BadgeIn × Skybridge moves authority outside language."

This positions the product as a solution to prompt injection, not just AI accuracy or control.

Inference: The positioning is rooted in security and trust, not general-purpose AI orchestration or productivity tools.

The description does not name specific customers or target industries.

It implies use cases in business systems where AI agents interact with sensitive data or actions (e.g., calendars, files, inboxes).

Inference: The system is likely aimed at enterprises or developers who manage AI agents in production and want to enforce human oversight for security-sensitive operations.

No evidence of pricing, monetization, or business model is provided.

The description focuses on the technical architecture and demo execution, not commercial viability.

Not evidenced

The system is built with:

• Python
• HTML/CSS/JavaScript
• Cloudflare Tunnel
• Qwen LLMs (local fallbacks)
• HMAC-authenticated grants
• Session sandboxed executor
• Pytest and adversarial test harnesses

It includes:

• Deterministic app manifests
• Same-origin human approval
• Hash-chained authority history
• Route failure narration (e.g., showing which fallback answered)

Inference: The system is built with a focus on security, determinism, and auditability. It uses local LLMs and sandboxed execution to reduce reliance on external APIs.

The description states:

• A demo was built during a hackathon (OpenAI 2026)
• The system passed 71 tests and 42 stress case groups
• It retained 0 unauthorized executions, detected all 3 evidence-tamper cases, preserved 3 authorized positive controls
• One winner among 48 concurrent attempts was allowed to spend a single-use grant

Not evidenced: No real-world usage, customers, revenue, or adoption data.

The description does not name competitors or reference existing tools in the market.

It positions itself as solving prompt injection and AI action control, which is a niche but emerging area of concern.

Inference: The product may be addressing a gap in AI governance and security, particularly around human-in-the-loop controls for AI agents.

• Demo-only: No evidence of production use or real-world integration.
• No commercial traction: No revenue, customers, or adoption data.
• Self-reported only: All claims are unverified.
• Limited team size: Only one member (Saul Lowery) is listed.
• Hackathon origin: The system was built in a short time for a contest, not as a long-term product.

1. What real-world use cases have you identified for this system beyond the demo?
2. Have you tested this with any external AI agents or systems in production?
3. How do you plan to scale human approval in high-throughput environments?
4. Are there any known limitations of the current sandboxed execution model?
5. What is your roadmap for moving from a hackathon demo to a production-ready product?