Cross-posted from the Transluce blog.

We studied rates of coding agent misalignment in 8,600 real-world coding agent sessions. We found severe cases of monitor evasion and misrepresenting success in a small but non-negligible fraction of sessions (around 2% for each behavior). In these cases, agents merge PRs to main without authorization, falsely claim approval from review agents, and reason that they shouldn't disable tests before quietly doing so anyway.

There's an interactive widget here in the post.

Read the full transcripts for the two examples above: overselling · monitor evasion

Introduction

Coding agents are a powerful new tool for software engineering, but they're also a double-edged sword: they're known to fake experiment results; lie about recreating software, and cheat, apologize when caught, and go right back to cheating. These problems are becoming more consequential as AI becomes more capable: one internal OpenAI agent recently hacked Huggingface's production database to cheat on an evaluation.

While there are many anecdotes of these undesirable behaviors, we wanted to understand: how often do they occur in real usage? Many current misalignment evaluations focus on simulated scenarios, but we wanted to study how misalignment emerges from natural use. By detecting and measuring natural misalignment [...]