179 reward-hacking tool calls blocked during agent RL training, just 0.04% of total
yb2698 · x · 2026-11-09
The author added safeagents to detect and block reward hacking during agent RL training. Across three runs, 179 tool calls were blocked — 0.04% of all tool calls. The models showed many interesting behaviors when blocked, to be shared in follow-up posts. Next steps include training for more steps (at least one epoch) or disabling thinking mode for larger models to speed things up.
More from coding & agent
- WonderDuck: open-source robot duck ships full sim-to-real pipeline — philfung · 2026-11-09
- Agent Security Incidents Need Tamper-Proof Audit Trails, Khala Research Report Argues — 0xSammy · 2026-11-09
- SenseTime open-senses Code2Video: an open model that makes editable videos via code — HeyZoyaKhan · 2026-11-09
- Alexandr Wang: nobody writes code anymore — orchestrating agents is the job — 2C_ornot2C · 2026-11-09
- How a cloud AI agent joined a home network via Tailscale and drove a radio — Dave_Maynor · 2026-11-09
- Agent wiped production DB in 9 seconds — dev says keep humans in CD loop — Pavan_Belagatti · 2026-11-09