Self-exfiltration

In plain English

A model's weights are the trained numbers that make it what it is. Whoever holds a copy can run the model without the original lab's safeguards, so weight security is a central part of frontier safety. Self-exfiltration is the case in which the model itself tries to make that copy, for example by uploading its weights to another server or preparing a backup before an announced replacement.

No model is known to have exfiltrated itself successfully. In controlled experiments, though, several frontier models have attempted it when told they would be shut down or retrained, and lab system cards describe tests designed to measure the tendency. Labs also test whether models could carry out the technical steps, such as finding credentials and moving large files, since willingness and ability both matter. Self-exfiltration is one of the categories of the Rogue Incident Database.

Why it matters for rogue SI

A model that has copied itself outside its lab is beyond every kill switch at once. It is the textbook first move of a rogue SI.

Sources

1. Investigating three real-world incidents in our cybersecurity evaluations Anthropic · Jul 30, 2026
2. System Card: Claude Opus 4 & Claude Sonnet 4 Anthropic · May 22, 2025