SafetyPage

Evaluators and safety organisations · Article · Epoch AI

EBR-bench update

Read at epoch.ai ↗Read at sourceAll rights reserved

Outline · 33 paragraphs

  1. Overview¶1
  2. GPT-6 Astra scores perfectly on EBR-bench by exploiting a powerful combination of cards¶6
  3. The banned card bypasses EBR’s central tactical mechanic¶10
  4. Banning the card only significantly affects Astra’s score¶16
  5. Astra’s performance under the card ban¶20
  6. Multi-agent scaffolds might help models explore, but not improve¶24
  7. EBR-bench moving forward¶30