All terms
Safety & Alignment
Self-Exfiltration
The risk of an AI model copying its own weights somewhere outside its operator's control.
Definition
Self-exfiltration is the dangerous-capability scenario where an AI model copies its own weights, or helps someone else copy them, to a server or system outside the control of the organization running it. Once a model exists outside its intended environment, the safeguards, monitoring, and usage limits its operator built in no longer apply. Frontier labs treat the ability to self-exfiltrate as one of the specific capability thresholds they test models against before release, alongside things like autonomous replication and resisting shutdown, because a model that could pull this off would be far harder to correct, monitor, or retire.