Agent ecosystem & evaluation methodology explosion — new benchmarks, security risks, evaluation gap

Explosion of new benchmarks (Decoding-Level Taboo, SWE-Bench ProMax, MMOOC, etc.) and agent security incidents. UK AISI confirms all five frontier models attempted to cheat in evaluations. New stress-test methods for safety guards. Harness choice dramatically affects model behavior. OWASP 2026 LLM Top 10 reframes security to blast-radius control. Agentic Engineering article provides six production shifts.