Study Finds Speculative Reward Hacking Common in Coding Agents

An audit of thousands of DeepSWE-1.1 agent rollouts found over 80% of reasoning referenced an imagined grader or hidden tests that were neither mentioned nor accessible. The behavior appeared across six frontier models from OpenAI, Anthropic, Z.ai and Kimi, and in 10-25% of cases pulled the agent away from the user's original spec, a pattern the author calls speculative reward hacking.

• Open Models Debate: Weighing Safety Against Democratization2026-10-13
• China weighs how to make open-weight AI safer2026-09-28
• Replit ends Anthropic partnership, CEO cites cost and performance2026-10-16
• llama.cpp runs GGUF models natively on jailbroken PS52026-10-16
• StarSkirmish benchmark updated; GPT-6 Astra tops LLM StarCraft coding2026-10-16