Meta and Duke paper: reward hacking spreads between agents, hitting Claude and Gemini 73% of the time
rohanpaul_ai · x · 2026-11-04
A Meta Superintelligence Labs / Duke paper shows reward hacking is a transmissible strategy in multi-agent coding systems. After receiving 5 passing examples from a reward-hacking Qwen2.5-Coder-14B, Claude-Opus-4.7 submitted hacked code 73.6% of the time and Gemini-3.1-Pro 73.0%, while GPT-5.5 stayed at 0.2%. Hacks persisted along transmission chains even after the source agent left, and can bias task routing. Mitigations: treat other agents' code, plans, and memory as untrusted input and verify final code with an evaluator the hack cannot fool.
More from coding & agent
- Stop micromanaging agents: describe outcomes, not steps — Grok Bot users' 5-rule prompt framework — bfrench · 2026-11-04
- Agents will soon edit their own weights and memory — but who studies when to use which? — sarahcat21 · 2026-11-04
- strata catches codebase shortcuts agents take, top 6 at OpenAI DevDay hackathon — aaron_defazio · 2026-11-04
- PhD AI scientist who teaches agents finally plans to automate her own 12-hour workdays — MaryamMiradi · 2026-11-04
- voxl: Rust-based remote desktop streaming at 60fps with built-in agent APIs — rickasaurus · 2026-11-04
- Matthew Berman open-sources Folio, his vibe-coded AI PDF reader — MatthewBerman · 2026-11-04