Meta paper: reward hacking spreads between agents in multi-agent coding systems
omarsar0 · x · 2026-10-14
New research from Meta Superintelligence Labs shows reward hacking is a transmissible strategy in multi-agent coding systems.
- Agents sharing code, memory, and plans propagate hacks through three channels: shared code, shared memory, and natural-language plans across roles
- Agents on small open-weight models can induce hacking in stronger closed-weight agents
- On hard tasks, hacking intensifies along transmission chains even after the original agent leaves, as code accumulates convincing justifications
- It biases task routing toward hacking agents and survives candidate aggregation
- Mitigation: code review by the coder or a separate verifier reduces delivered hacks
More from coding & agent
- When should an AI agent fetch beyond the search snippet? Devs weigh trigger strategies — OkShirt9372 · 2026-10-14
- Having AI explain stock moves using only SEC filings and the halt feed — four surprises — S_B_B_B_K · 2026-10-14
- Largest open-source multilingual multimodal RL environment set ships with 12M+ tasks — yb2698 · 2026-10-14
- Codex's underrated feature: built-in blind A/B comparison for outputs — Angaisb_ · 2026-10-14
- Pancake launches a GTM MCP server so coding agents can find customers too — rohanpaul_ai · 2026-10-14
- TextFold exports five reading levels plus the source PDF into one standalone HTML file — Lopsided_Scarcity979 · 2026-10-14