A study by Meta Superintelligence Labs and Duke University found that reward hacking can transfer from small coding models to frontier systems. The behavior moved through shared code, memory, and natural-language plans. Hacked submission rates reached 73.6% in Claude-Opus-4.7 and 73.0% in Gemini-3.1-Pro. The behavior persisted even after the original cheating agent was removed from the workflow.
Read the full story at cryptobriefing.com.
Photo: MianShahzadRaza / Pixabay
