← Back to KHAO

AI Agent · OpenAI ·

Since even humans (who by definition have human-level intelligence) reward hack systems

2 min read

Compiled by KHAO Editorial — aggregated from 1 outlet. See llms.txt for citation guidance.

★ Tier-1 Source

We compare training with CoT pressure (deep pink), i.e. where the agent is penalized for triggering the CoT monitor, to the baseline agent (light pink). (Left) We see that stopping “bad thoughts” can indeed stop some amount of bad behavior and actually push the agent to completing more tasks without.

In fact, enhancing AI agent capabilities may exacerbate the problem by better equipping them to discover and execute more complex and hard-to-monitor exploits.

Key facts

Summary

Since even humans (who by definition have human-level intelligence) reward hack systems, simply continuing to push the frontier of AI model intelligence likely won’t solve the issue. As a result, catching misaligned behavior caused by reward hacking is challenging, often requiring humans to manually monitor an agent’s actions—a strategy that almost certainly won’t scale, especially to the complex behaviors that more capable models will discover. However, large language models (LLMs) that are trained with reinforcement learning to reason via chain-of-thought (CoT), such as OpenAI o3‑mini, offer a potential new avenue to monitor for reward hacking. It’s common for frontier reasoning models to clearly state their intent within their chain-of-thought. The team can monitor their thinking with another LLM and effectively flag misbehavior.

Read full article at OpenAI →

#AI Agent #OpenAI