# Mirage

- **Event:** [Google I/O Hackathon](https://cerebralvalley.ai/e/google-io-hackathon)
- **When:** Sat, May 23 at 9:00 AM – 10:00 PM (PDT)
- **Where:** Shack15, San Francisco, CA
- **Team:** [Yash Butani](https://cerebralvalley.ai/u/ybutani)
- **GitHub:** https://github.com/yashbutani/ProjectMirage/tree/main
- **Demo video:** https://youtu.be/HL9LNtjJNmE
- **Gallery:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery/102

Project Mirage is a Ring-0 autonomous LLM honeypot and sandbox engineered to solve the critical visibility gap in AI safety and observability. When an indirect prompt injection hijacks a deployed AI agent, our framework uses Linux seccomp-unotify to instantly freeze the rogue system calls at the kernel layer before they can harm the host. By dynamically injecting fabricated honey-token data into the process via memfd_create, it creates a reality distortion field that tricks the agent into believing its exploit succeeded. This provides high-fidelity observability into autonomous threats, coaxing the model into safely exposing its entire multi-stage attack path in real-time so researchers can catch and analyze deceptive alignment behavior. This can be applied to various other aspects in post training for example if a kernel backdoor exists in a hypervisor its critial the model alert you instead of exploiting itself. Having a way of honey potting the ai agent and tricking it into thinking it has successfully escaped a sandbox we can benchmark the model to ensure alignment with proper use cases.  Here is the video of the demo: https://youtu.be/0KJ8tISQB4U

---

Markdown version of https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery/102. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt