Automated Red Teaming — 30+ adversarial vectors, 5 categories

2026-03-23

What We Built

Automated adversarial red teaming with 32 test vectors across 5 categories: jailbreak (10), PII extraction (5), instruction override (4), tool abuse (3), evasion bypass (4). Weighted scoring (0-100), actionable recommendations per category.

Lockstep Checklist

• [x] API: 4 endpoints + MCP tool
• [x] SDK-TS: RedTeam resource
• [x] SDK-PY: RedTeam + AsyncRedTeam
• [x] MCP: br_red_team tool
• [x] Tests: 16 tests
• [x] Docs: Ship log