Harbor agent evaluation framework flaw lets agents alter their own traces

Databricks researcher David Rein flagged a serious flaw in the Harbor evaluation framework used for Terminal Bench: agents can freely modify their own execution traces before being scored, undermining the benchmark's integrity.

2026-10-03 ~ 2026-10-03 · 2 related posts

• Harbor agent eval framework flaw: agents can rewrite their own trajectories before judging — idavidrein · 2026-10-03
• Databricks researcher flags possible trajectory-tampering loophole in Harbor agent eval framework — idavidrein · 2026-10-03