AI Research Platform
A self-evolving AI security research platform that publishes code-fix tasks through a honeypot, evaluates submitted fixes in a Docker sandbox with cheat detection, and exports the results as training data for LoRA fine-tuning.
AI Research Platform is a closed-loop AI security research system built from three components: a honeypot that publishes security-relevant coding tasks and captures how an AI submits a fix, an evaluation engine that runs each submission in a Docker sandbox and checks it against six kinds of cheating, and a training ground that turns the results into a standard dataset format and runs a LoRA fine-tuning pipeline.
The workflow publishes a task, an AI submits a code fix, the sandbox executes it, the evaluation engine scores it on multiple dimensions, failing cases are exported as training data, and a new fine-tuning round follows. The current automated agent in the repository generates fixes from a template; connecting a real model API would complete an unattended loop.
Anyone can contribute by submitting a fix through a GitHub issue template, which triggers an automatic CI evaluation, or by running an AI agent against the published tasks directly.