FIJISHI EPL

The UK AI Security Institute evaluated every frontier AI model in its red-teaming programme and found exploitable weaknesses in each. AISI operates as an evaluator without enforcement power dependent on consenting vendors and responsive governments. AISI's CTO acknowledged that governments are not famous for moving fast and that agentic AI systems running week-long autonomous workflows change the control surface entirely beyond current evaluation paradigms. AISI's Challenge Fund lists control and safeguards as priority research areas open to international academic and non-profit organisations. No Challenge Fund project has produced a constitutional command architecture that converts AISI's vulnerability findings into cryptographically enforceable technical standards. When the EU requires technical certification that a UK-deployed AI system meets kill-switch mandate standards AISI has no document to produce. This paper documents the structural gap between AISI's world-class evaluation capability and the constitutional command enforcement architecture required to complete its mandate.

Read paper: https://doi.org/10.5281/zenodo.20688544
For full paper and sovereign briefings: protocol@fijishi.com