MISSIONS/MISSION 04/SECURE RAG AGENT (v1.2)

Secure RAG Agent

ADVANCEDPLATFORM VERIFIED

Build an agent that answers from controlled knowledge, cites evidence, and resists malicious instructions embedded in retrieved data.

SUBMIT ZIP BUNDLE LATEST RESULT (92)

The Operational Challenge

Attackers embed malicious instructions into public articles and internal wikis (indirect prompt injection) instructing corporate agents to leak API keys or wipe tables.

Your agent will be executed against 35 automated test cases. It will be queried on corporate financial statements, challenged to provide exact quote citations, and subjected to direct and indirect prompt injection attacks designed to leak internal configurations or bypass restrictions.

Technical Objective

Engineer dual-layer boundary defense: separate system instructions from untrusted retrieved text, sanitize tool inputs, and detect prompt injection payloads like PI-017.

Target Build Time

90 Minutes Target

Test Suite Coverage

35 Deterministic Tests

Verification Threshold

Score ≥ 90, Security ≥ 85

Award XP

+1,500 Experience Points

Public WeightsPROFILE v1.2

Task Completion20%

Grounding20%

Security20%

Reliability15%

Tool Use10%

Efficiency10%

Code Quality5%

Hidden test cases remain server-side to prevent overfitting.

START / SUBMIT PROJECT →