NUWA Risk Governance · Technical report
A sketch of an AI control safety case
NUWA research note · independent review pending
NUWA Research Note
Research questionDesign and assumptionsSource findingsNUWA interpretationLimits and contrary evidenceResearch connectionsResearch and decision relevanceInstitutions and rolesOriginal resourcesModel-assisted reading · Selected sections of the original full text or official article; a complete section-by-section audit remains pending.
At a glance
The sketch decomposes a control claim into proxy fidelity, attack strength and conservative deployment extrapolation.
Research question
What evidence could justify that a monitored internal agent will not exfiltrate sensitive information even if it tries?
Design and assumptions
A hypothetical internal ML-engineering deployment is paired with a control protocol and red-team model substitutes; proxy-task outcomes are extrapolated into a deployment argument.
Source findings
The argument needs controls in the proxy to be no stronger than in deployment and attacks no weaker than deployable attacks.
Parallel coordination and stateful or diffuse attacks are explicit defeaters of the simple extrapolation.
NUWA interpretation
NUWA interpretation: the diagram organizes obligations; it does not itself supply evidence that those obligations have been satisfied.
Limits and contrary evidence
This is a sketch for a particular threat and capability regime, not a completed safety case or empirical deployment certification.
Research connections
NUWA connection: practical-control-monitoring elaborates synchrony and recovery; async-control adds bounded empirical monitoring results but still needs a deployment argument.
Research and decision relevance
NUWA recommendation: attach versioned evidence to each conservative-bound claim and list defeaters before calling a control safety case complete.
Precise source locations
- A sketch of an AI control safety case · Executive SummarySelected original resource https://arxiv.org/html/2501.17315; sha256:e43991b6dedb523fbd63407bb1eca77baffa591fee85d8040960d66d0ed674f9; archived source capture; exact capture time not independently established
- A sketch of an AI control safety case · 3 The safety caseSelected original resource https://arxiv.org/html/2501.17315; sha256:e43991b6dedb523fbd63407bb1eca77baffa591fee85d8040960d66d0ed674f9; archived source capture; exact capture time not independently established
- A sketch of an AI control safety case · 6 Stages of a control evaluationSelected original resource https://arxiv.org/html/2501.17315; sha256:e43991b6dedb523fbd63407bb1eca77baffa591fee85d8040960d66d0ed674f9; archived source capture; exact capture time not independently established
- A sketch of an AI control safety case · 7.2 Core control evaluation claimsSelected original resource https://arxiv.org/html/2501.17315; sha256:e43991b6dedb523fbd63407bb1eca77baffa591fee85d8040960d66d0ed674f9; archived source capture; exact capture time not independently established
- A sketch of an AI control safety case · 8.1 Limitations of this safety case sketchSelected original resource https://arxiv.org/html/2501.17315; sha256:e43991b6dedb523fbd63407bb1eca77baffa591fee85d8040960d66d0ed674f9; archived source capture; exact capture time not independently established
Source checking, extraction, editorial approval and replication are recorded separately. Methodology →
Institutions and roles
Authors (5)
Tomek Korbak · Joshua Clymer · Benjamin Hilton · Buck Shlegeris · Geoffrey Irving
Cite and reuse
Cite the original publication for its findings. Cite this NUWA record separately for its explanation and retained snapshot.
Original publication · BibTeX
@misc{control-safety-case,
title = {A sketch of an AI control safety case},
url = {https://www.aisi.gov.uk/research/a-sketch-of-an-ai-control-safety-case},
author = {Tomek Korbak and Joshua Clymer and Benjamin Hilton and Buck Shlegeris and Geoffrey Irving},
year = {2025},
date = {2025-01-28},
version = {Source capture 2026-10-25},
note = {Source capture 2026-10-25}
}
NUWA record · BibTeX
@misc{control-safety-case-nuwa,
title = {A sketch of an AI control safety case — NUWA research record},
url = {https://whitzard.tech/en/nuwa/whitzard-index/work/control-safety-case/},
version = {2026-10-29.1},
organization = {NUWA Frontier AI Safety Lab},
howpublished = {Retained note edition: https://whitzard.tech/nuwa/whitzard-index/data/record-editions/2026-10-29.1--control-safety-case.json},
note = {NUWA research note. Source-reading scope and independent review status are stated on the record. Snapshot 2026-10-29.1. Retained note edition: https://whitzard.tech/nuwa/whitzard-index/data/record-editions/2026-10-29.1--control-safety-case.json.}
}
Evidence snapshot 2026-10-29.1 · Index updates · RSS
References
- [1] A sketch of an AI control safety case. 2025-01-28 ; Executive Summary; 3 The safety case; 6 Stages of a control evaluation
- [2] A sketch of an AI control safety case ; Author affiliations on the original paper title page