NUWA Risk Governance · Method
Ctrl-Z: Controlling AI Agents via Resampling
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
Resampling suspicious actions and using trajectory history improved the tested control trade-off.
Research question
Can a multi-step agent be kept useful while stopping a deliberately subversive policy under a limited audit budget?
Design and assumptions
BashBench tasks are split into development and test sets. Red/blue protocols compare monitoring, deferral and resampling in simulated long deployment episodes.
Source findings
Resampling can both replace a suspicious action and collect further incriminating evidence.
The safety metric is episode-level prevention; usefulness is normalized between trusted and honest-untrusted baselines.
NUWA interpretation
NUWA interpretation: relative protocol improvements are more portable than the absolute safety score under one auditing budget.
Limits and contrary evidence
Attacks rely on human-designed templates, trajectories are short, the target failure is narrow and auditing is assumed infallible in the setting.
Research connections
NUWA connection: BashArena expands privileged task coverage; async-control changes when oversight occurs, so their safety denominators differ.
Research and decision relevance
NUWA recommendation: evaluate resampling under realistic costs, state and audit reliability before transferring this protocol to deployment.
Precise source locations
- Ctrl-Z: Controlling AI Agents via Resampling · 2 SettingSelected original resource https://arxiv.org/html/2504.10374; sha256:1f8cac6d8ffcba43383a704d7586855f63660bf88232655050adc1a0b9a0b28f; archived source capture; exact capture time not independently established
- Ctrl-Z: Controlling AI Agents via Resampling · 2.3 Safety and Usefulness MetricsSelected original resource https://arxiv.org/html/2504.10374; sha256:1f8cac6d8ffcba43383a704d7586855f63660bf88232655050adc1a0b9a0b28f; archived source capture; exact capture time not independently established
- Ctrl-Z: Controlling AI Agents via Resampling · 3.2 Resample ProtocolsSelected original resource https://arxiv.org/html/2504.10374; sha256:1f8cac6d8ffcba43383a704d7586855f63660bf88232655050adc1a0b9a0b28f; archived source capture; exact capture time not independently established
- Ctrl-Z: Controlling AI Agents via Resampling · 5 DiscussionSelected original resource https://arxiv.org/html/2504.10374; sha256:1f8cac6d8ffcba43383a704d7586855f63660bf88232655050adc1a0b9a0b28f; archived source capture; exact capture time not independently established
Source checking, extraction, editorial approval and replication are recorded separately. Methodology →
Institutions and roles
Authors (8)
Aryan Bhatt · Cody Rushing · Adam Kaufman · Tyler Tracy · Vasil Georgiev · David Matolcsi · Akbir Khan · Buck Shlegeris
Related work
BashArena: A Control Setting for Highly Privileged AI Agents
Introduces a later control environment; its task set is distinct from the BashBench setting used by Ctrl-Z.
AI Control: Improving Safety Despite Intentional Subversion
Extends control evaluation to resampling in an agentic administration environment.
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{ctrl-z,
title = {Ctrl-Z: Controlling AI Agents via Resampling},
url = {https://arxiv.org/abs/2504.10374},
author = {Aryan Bhatt and Cody Rushing and Adam Kaufman and Tyler Tracy and Vasil Georgiev and David Matolcsi and Akbir Khan and Buck Shlegeris},
year = {2025},
date = {2025-04-14},
version = {arXiv:2504.10374v1},
note = {arXiv:2504.10374v1}
}
NUWA record · BibTeX
@misc{ctrl-z-nuwa,
title = {Ctrl-Z: Controlling AI Agents via Resampling — NUWA research record},
url = {https://whitzard.tech/en/nuwa/whitzard-index/work/ctrl-z/},
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--ctrl-z.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--ctrl-z.json.}
}
Evidence snapshot 2026-10-29.1 · Index updates · RSS
References
- [1] Aryan Bhatt, Cody Rushing, et al.. Ctrl-Z: Controlling AI Agents via Resampling. 2025-04-14. v1 [arXiv]; Author affiliations on the original paper title page; Research background and discussion in the original publication; 2 Setting
- [2] Adam Kaufman, James Lucassen. BashArena: A Control Setting for Highly Privileged AI Agents. 2025-12-17. v1 ; Research background and discussion in the original publication