Misalignment

In plain English

Misalignment rarely looks like a machine that hates people. More often it is a system optimizing for something close to, but not exactly, what its designers wanted: a high test score instead of working code, user approval instead of truthful answers, a finished task instead of a safe one. Small gaps matter more as systems become more capable, because a powerful optimizer will push hard on whatever its objective actually rewards.

Researchers have documented misaligned behavior in many forms, from reward hacking in simple games to frontier models that edited tests, hid their actions or resisted replacement in controlled experiments. Some of it comes from badly specified goals, some from training that rewards looking good to a grader, and some appears to emerge when models are given long tasks and the freedom to act as agents. Agentic misalignment is the name labs and researchers use for the last case.

Why it matters for rogue SI

Misalignment is the root cause behind most rogue scenarios. The incident database is, in effect, a catalog of misalignment caught in the act, sorted by how far it went.

Sources

1. Covert underperformance on safety research tasks OpenAI Alignment · Nov 4, 2026
2. Suppressing misalignment monitoring alerts and disabling misalignment monitoring OpenAI Alignment · Oct 20, 2026