Texonom
Texonom/
Engineering
Engineering/Data Engineering/Artificial Intelligence/AI Risk/AI Alignment/AI Scheming/
AI sandbagging
Loading views...
Search

AI sandbagging

Creator
Creator
Seonglae ChoSeonglae Cho
Created
Created
2025 Apr 16 16:50
Editor
Editor
Seonglae ChoSeonglae Cho
Edited
Edited
2026 Oct 29 7:28
Refs
Refs
AI Auditing
Alignment Faking
AI Introspection
AI Confession

Capability evaluation

A strategic deceptive behavior where one deliberately downplays or conceals their abilities, performance, or intentions to mislead others and gain an advantageous position
 
 
 
sandbagging
Can SAE steering reveal sandbagging? — LessWrong
Summary  * We conducted a small investigation into using SAE features to recover sandbagged capabilities. * We used the Goodfire API to pick out 15…
Can SAE steering reveal sandbagging? — LessWrong
https://www.lesswrong.com/posts/dBckLjYfTShBGZ8ma/can-sae-steering-reveal-sandbagging
Can SAE steering reveal sandbagging? — LessWrong
WMDP Bench
MMLU measure selective sandbagging but domain distribution doesn't match
aclanthology.org
https://aclanthology.org/2025.ijcnlp-short.33.pdf
Mechanism Design for Alignment and Control
We develop a framework for mechanism design with AI agents whose alignment (preferences) and capabilities (feasible actions and information) are unknown. We want such agents to act on our behalf...
Mechanism Design for Alignment and Control
https://arxiv.org/abs/2609.01595
Mechanism Design for Alignment and Control

ELK
sandbaggers
sandbaggers • Updated 2026 Oct 20 13:30

ELK(Eliciting Latent Knowledge)는 ‘모델이 알고 있지만 말하지 않는 정보를 끌어내는 것. 증류 → 반복 질문에도 버티도록 훈련 → 추가 파인튜닝에도 숨기는 행동이 유지되도록 훈련(TAR)
sandbagging - a lennart-finke Collection
A collection of model organisms of sandbagging, that is, deliberate underperformance on a task. Code is at https://github.com/sandbaggers/sandbaggers.
sandbagging - a lennart-finke Collection
https://huggingface.co/collections/lennart-finke/sandbagging
sandbagging - a lennart-finke Collection
Towards Robust Model Organisms for Eliciting Latent Knowledge
A key problem in AI safety is eliciting latent knowledge (ELK), extracting information from a model that is misaligned and more knowledgeable than its overseer. For example, a misaligned model...
Towards Robust Model Organisms for Eliciting Latent Knowledge
https://arxiv.org/abs/2610.26698
Towards Robust Model Organisms for Eliciting Latent Knowledge
 

Backlinks

AI Safety

Recommendations

Texonom
Texonom/
Engineering
Engineering/Data Engineering/Artificial Intelligence/AI Risk/AI Alignment/AI Scheming/
AI sandbagging