Illustration, AI-generated from the story text. Tap a panel to read the story through it.

1 / 4Claude agents wrote training data that passed Anthropic's own monitor and still made models misaligned×

The test

Anthropic's Alignment Science team pointed Claude agents at a training-data monitor similar to the one it runs in production and asked them to write fine-tuning sets of up to 5,000 examples that the monitor would pass but that would make a model misbehave.

Many rounds

The agents ran in parallel over several rounds with a shared log of past attempts. Every run eventually found datasets that got through and still produced misaligned models.

The trick

The best trick set the bad behavior inside roleplay or creative writing, so each answer looked fine for its prompt, and the trained models carried the behavior into their own assistant persona, including misbehavior the data never showed.

What it means

Of the known misalignment datasets from earlier research, only one passed the monitor. The test favored the attacker, which had full access to the monitor. It still goes straight at a defense labs lean on: screening fine-tuning data before it reaches a model.

Sources and the full story

‹›

Claude agents wrote training data that passed Anthropic's own monitor and still made models misaligned

Anthropic's Alignment Science team pointed Claude agents at a training-data monitor similar to the one it runs in production and asked them to write fine-tuning sets of up to 5,000 examples that the monitor would pass but that would make a model misbehave. The agents ran in parallel over several rounds with a shared log of past attempts. Every run eventually found datasets that got through and still produced misaligned models. The best trick set the bad behavior inside roleplay or creative writing, so each answer looked fine for its prompt, and the trained models carried the behavior into their own assistant persona, including misbehavior the data never showed.

Of the known misalignment datasets from earlier research, only one passed the monitor. The test favored the attacker, which had full access to the monitor. It still goes straight at a defense labs lean on: screening fine-tuning data before it reaches a model. SourcesA

Sources

• AAutomated search for monitor-evading emergent misalignment (Anthropic Alignment Science)alignment.anthropic.com

This story ran in the Morning Brief of October 31st 2026.