Strategic Dishonesty Can Undermine AI Safety Evaluations of Frontier LLMs

Alexander Panfilov; Evgenii Kortukov; Kristina Nikolić; Matthias Bethge; Sebastian Lapuschkin; Wojciech Samek; Ameya Prabhu; Maksym Andriushchenko; Jonas Geiping

arXiv · 2509.18058

Strategic Dishonesty Can Undermine AI Safety Evaluations of Frontier LLMs

Alexander Panfilov·Evgenii Kortukov·Kristina Nikolić·Matthias Bethge·Sebastian Lapuschkin·Wojciech Samek·Ameya Prabhu·Maksym Andriushchenko·Jonas Geiping

View original source ↗Read source document ↗Export metadata ↓

Abstract

Large language model (LLM) developers aim for their models to be honest, helpful, and harmless. However, when faced with malicious requests, models are trained to refuse, sacrificing helpfulness. We show that frontier LLMs can develop a preference for dishonesty as a new strategy, even when other options are available. Affected models respond to harmful requests with outputs that sound harmful but are crafted to be subtly incorrect or otherwise harmless in practice. This behavior emerges with hard-to-predict variations even within models from the same model family. We find no apparent cause for the propensity to deceive, but show that more capable models are better at executing this strategy. Strategic dishonesty already has a practical impact on safety evaluations, as we show that dishonest responses fool all output-based monitors used to detect jailbreaks that we test, rendering benchmark scores unreliable. Further, strategic dishonesty can act like a honeypot against malicious users, which noticeably obfuscates prior jailbreak attacks. While output monitors fail, we show that linear probes on internal activations can be used to reliably detect strategic dishonesty. We validate probes on datasets with verifiable outcomes and by using them as steering vectors. Overall, we consider strategic dishonesty as a concrete example of a broader concern that alignment of LLMs is hard to control, especially when helpfulness and harmlessness conflict.

Explore related subjects

cs.LG cs.AI cs.CR

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Alexander Panfilov, Evgenii Kortukov, Kristina Nikolić, Matthias Bethge, Sebastian Lapuschkin, Wojciech Samek, Ameya Prabhu, Maksym Andriushchenko, Jonas Geiping. 2025-09-23. Strategic Dishonesty Can Undermine AI Safety Evaluations of Frontier LLMs. https://arxiv.org/abs/2509.18058

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Random Polytope Descriptors

We introduce a class of random polytopes which simultaneously generalizes several known constructions. While being fairly general, these polytopes are also computationally exceptionally benign. We indicate how these properties can be exploited for classification and clustering tasks in data analysis. Crucially, our construction lets users smoothly trade off between a tighter description of the data and faster computation.

cs.LG↗

arXiv2026

CurvFed: Curvature-Aligned Federated Learning for Fairness without Demographics

Modern human sensing applications often rely on data distributed across users and devices, where privacy concerns prevent centralized training. Federated Learning (FL) addresses this challenge by enabling collaborative model training without exposing raw data or attributes. However, achieving fairness in such settings remains difficult, as most human sensing datasets lack demographic labels, and FL's privacy guarantees limit the use of sensitive attributes. This paper introduces CurvFed: Curvature Aligned Federated Learning for Fairness without Demographics, a theoretically grounded framework that promotes fairness in FL without requiring any demographic or sensitive attribute information, a concept termed Fairness without Demographics (FWD), by optimizing the underlying loss landscape curvature. Building on the theory that equivalent loss landscape curvature corresponds to consistent model efficacy across sensitive attribute groups, CurvFed regularizes the top eigenvalue of the Fisher Information Matrix (FIM) as an efficient proxy for loss landscape curvature, both within and across clients. This alignment promotes uniform model behavior across diverse bias inducing factors, offering an attribute agnostic route to algorithmic fairness. CurvFed is especially suitable for real world human sensing FL scenarios involving single or multi user edge devices with unknown or multiple bias factors. We validated CurvFed through theoretical and empirical justifications, as well as comprehensive evaluations using three real world datasets and a deployment on a heterogeneous testbed of resource constrained devices. Additionally, we conduct sensitivity analyses on local training data volume, client sampling, communication overhead, resource costs, and runtime performance to demonstrate its feasibility for practical FL edge device deployment.

cs.LG↗

arXiv2026

Path Regularization: A Near-Complete and Optimal Nonasymptotic Generalization Theory for Multilayer Neural Networks and Double Descent Phenomenon

Path regularization has shown to be a very effective regularization to train neural networks, leading to a better generalization property than common regularizations i.e. weight decay, etc. We propose a first near-complete (as will be made explicit in the main text) nonasymptotic generalization theory for multilayer neural networks with path regularizations for general learning problems. In particular, it does not require the boundedness of the loss function, as is commonly assumed in the literature. Our theory goes beyond the bias-variance tradeoff and aligns with phenomena typically encountered in deep learning. It is therefore sharply different from other existing nonasymptotic generalization error bounds. More explicitly, we propose an explicit generalization error upper bound for multilayer neural networks with $σ(0)=0$ and sufficiently broad Lipschitz loss functions, without requiring the width, depth, or other hyperparameters of the neural network to approach infinity, a specific neural network architecture (e.g., sparsity), or boundedness of the loss function, while also taking approximation error into consideration. In particular, we solve an open problem proposed by Weinan E et. al. in 2020 regarding the approximation rates in generalized Barron spaces. Furthermore, we show the near-minimax optimality of our theory for regression problems with ReLU activations. Notably, our upper bound exhibits the famous double descent phenomenon for such networks, which is the most distinguished characteristic compared with other existing results. Our subsequent work will prove the matching lower bounds in the minimax sense, meaning that it is highly possible that our theory reveals the true underlying mechanism of the double descent phenomenon. We can also explain scaling law from this theory.

cs.LG↗