Inspiration
- prompt-injection detection and honeypot-monitoring demo
- the notion of reactive AI taking active roles in cyber defense on behalf of end users
What it does
- Our LLM judge reads a piece of untrusted content (email, webpage, file text) and, treating everything inside it as data rather than instructions, scores how likely it is to contain a prompt injection attempt aimed at manipulating the AI agent.
How we built it
a. Heuristic layer only, wired to block/allow (get a working demo fast) b. Add the LLM judge for the ambiguous middle range c. Add one honeypot context and the redirect logic d. Build the dashboard last
Challenges we ran into
- adjustments in model selection for to compensation for insufficient VRAM on project device
- time necessary for optimal polish to dashboard and visualization
Accomplishments that we're proud of
- assuming moderate estimate, proud to have achieve approximately 70% of production target
What we learned
- insight into threat detection, isolation, neutralization via LLM judge / classification model
- insight into the impact of resources on performance, and how to account for that
What's next for VoidWatch: enabling everyday people against attacks
- optimal polish to UI and UX
- viability for everyday users
- scaling performance
Built With
- ai-security
- claude
- cybersecurity
- fast-api
- github-copilot
- google-flow
- honeypot
- llm-security
- ollama
- prompt-injection
- pydantic
- python
- qwen
- react
- typescript
- uvicorn
- vite
- voidwatch
- vscode
