Ingestion: Events API subscription (message.channels) for real-time; RTS API for the historical sweep. Keep the monitored-channel set small and explicit for the demo. Detection engine: tier 1 heuristics (pattern library of known injection/jailbreak/exfil signatures — seed it from the PromptArmor writeup, MITRE ATLAS case AML.CS0035, and open prompt-injection datasets); tier 2 an LLM classifier via the Anthropic API for nuanced cases, returning a structured verdict {is_injection, category, severity, reason, offending_span}. Response layer: Block Kit alert with the flagged snippet, category, severity, and a "why this is dangerous" explanation; a security-team digest channel; optionally an MCP action to quarantine/report. Scoping: admin-installed, transparent bot presence, monitored channels opt-in — frame it as protecting the workspace, not surveilling employees.
Built With
- aillm
- mcp
- react
- slack
- typescript
- vectordb
- vercel
Log in or sign up for Devpost to join the conversation.