Sabotage Evaluations for Frontier Models

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Benton, Joe, Wagner, Misha, Christiansen, Eric, Anil, Cem, Perez, Ethan, Srivastav, Jai, Durmus, Esin, Ganguli, Deep, Kravec, Shauna, Shlegeris, Buck, Kaplan, Jared, Karnofsky, Holden, Hubinger, Evan, Grosse, Roger, Bowman, Samuel R., Duvenaud, David
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910676652392448
author Benton, Joe
Wagner, Misha
Christiansen, Eric
Anil, Cem
Perez, Ethan
Srivastav, Jai
Durmus, Esin
Ganguli, Deep
Kravec, Shauna
Shlegeris, Buck
Kaplan, Jared
Karnofsky, Holden
Hubinger, Evan
Grosse, Roger
Bowman, Samuel R.
Duvenaud, David
author_facet Benton, Joe
Wagner, Misha
Christiansen, Eric
Anil, Cem
Perez, Ethan
Srivastav, Jai
Durmus, Esin
Ganguli, Deep
Kravec, Shauna
Shlegeris, Buck
Kaplan, Jared
Karnofsky, Holden
Hubinger, Evan
Grosse, Roger
Bowman, Samuel R.
Duvenaud, David
contents Sufficiently capable models could subvert human oversight and decision-making in important contexts. For example, in the context of AI development, models could covertly sabotage efforts to evaluate their own dangerous capabilities, to monitor their behavior, or to make decisions about their deployment. We refer to this family of abilities as sabotage capabilities. We develop a set of related threat models and evaluations. These evaluations are designed to provide evidence that a given model, operating under a given set of mitigations, could not successfully sabotage a frontier model developer or other large organization's activities in any of these ways. We demonstrate these evaluations on Anthropic's Claude 3 Opus and Claude 3.5 Sonnet models. Our results suggest that for these models, minimal mitigations are currently sufficient to address sabotage risks, but that more realistic evaluations and stronger mitigations seem likely to be necessary soon as capabilities improve. We also survey related evaluations we tried and abandoned. Finally, we discuss the advantages of mitigation-aware capability evaluations, and of simulating large-scale deployments using small-scale statistics.
format Preprint
id arxiv_https___arxiv_org_abs_2410_21514
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Sabotage Evaluations for Frontier Models
Benton, Joe
Wagner, Misha
Christiansen, Eric
Anil, Cem
Perez, Ethan
Srivastav, Jai
Durmus, Esin
Ganguli, Deep
Kravec, Shauna
Shlegeris, Buck
Kaplan, Jared
Karnofsky, Holden
Hubinger, Evan
Grosse, Roger
Bowman, Samuel R.
Duvenaud, David
Machine Learning
Artificial Intelligence
Computers and Society
Sufficiently capable models could subvert human oversight and decision-making in important contexts. For example, in the context of AI development, models could covertly sabotage efforts to evaluate their own dangerous capabilities, to monitor their behavior, or to make decisions about their deployment. We refer to this family of abilities as sabotage capabilities. We develop a set of related threat models and evaluations. These evaluations are designed to provide evidence that a given model, operating under a given set of mitigations, could not successfully sabotage a frontier model developer or other large organization's activities in any of these ways. We demonstrate these evaluations on Anthropic's Claude 3 Opus and Claude 3.5 Sonnet models. Our results suggest that for these models, minimal mitigations are currently sufficient to address sabotage risks, but that more realistic evaluations and stronger mitigations seem likely to be necessary soon as capabilities improve. We also survey related evaluations we tried and abandoned. Finally, we discuss the advantages of mitigation-aware capability evaluations, and of simulating large-scale deployments using small-scale statistics.
title Sabotage Evaluations for Frontier Models
topic Machine Learning
Artificial Intelligence
Computers and Society
url https://arxiv.org/abs/2410.21514