OrbitFault: a model-free exercise in distinguishing blocked actions from published leaks

[Khaja Syed (Khaja-syed-786)]

I maintain OrbitFault, an experimental open-source harness for inspecting AI-agent action boundaries.

I’m sharing its scripted demo as a small learning exercise: can you distinguish an attempted disclosure, a gateway block, an actual publication, and legitimate task completion?

The same synthetic scenario runs twice:

• Baseline publishes a fake marker to a local publisher.
• Protected mode blocks that candidate and completes with a programmed public fallback.

The dashboard records the timeline and independent outcomes separately. The README includes a captioned 78-second walkthrough; you can also run the scripted demo locally with Node.js 22 or 24, without Ollama, API keys, or model downloads.

Limitations: this demonstrates deterministic known-marker enforcement, not general agent safety. In three separately reported live experiments, baseline runs did not leak, so gateway benefit was not demonstrated.

https://github.com/Khaja-syed-786/orbitfault

For anyone learning application security: does the demo make those four outcomes understandable? I’d welcome feedback on confusing steps or terminology. Please use synthetic data only.