PROJECT_DG_01 / Security
MCP Sentinel
////// Summary
A security gateway prototype for AI tool calls, with attack scenarios, provenance traces, and allow / ask / block decisions.
- Category
- Security
- Status
- Prototype
- Contribution
- Security engineering · team project
- Stack
- Python · MCP · React · TypeScript · WebSocket
- Date
- 2026
/// My contribution
Attack-lab scenarios and security-engine work. The dashboard was developed with teammates; its default demonstration uses simulated events.
/// The boundary
An agent can encounter hostile instructions in a tool response. Those instructions should not acquire the authority of the user. Sentinel explores enforcing that boundary before a proposed action executes.
/// Attack lab
The project work includes prompt injection, cross-tool escalation, tool poisoning, and a tool-definition rug pull. The scenarios distinguish the origin of an instruction from the capability the agent tries to exercise.
/// An inspectable decision
The frontend separates the event stream, decision details, configuration, and history. A shared event contract connects the visual trace to the gateway interface. Default mock mode allows the team to rehearse scenarios independently of backend availability.
/// Current evidence
The public frontend and its simulated decision flow are inspectable. End-to-end enforcement against a connected live gateway has not been verified for this portfolio. The demonstration interface is not evidence of production protection.
/// System flow
- 01User intent
- 02Agent tool call
- 03Gateway checks
- 04Allow / Ask / Block