Machine learning models are routinely deployed on a wide range of computing hardware. Although such hardware is typically expected to produce identical results, differences in its design can lead to small numerical variations during inference. In this work, we show that these variations can be exploited to create backdoors in machine learning models. The core idea is to shape the model's decision function such that it yields different predictions for the same input when executed on different hardware. This effect is achieved by locally moving the decision boundary close to a target input and then refining numerical deviations to flip the prediction on selected hardware. We empirically demonstrate that these hardware-triggered backdoors can be created reliably across common GPU accelerators. Our findings reveal a novel attack vector affecting the use of third-party models, and we investigate different defenses to counter this threat.

In distributed applications, like swarms of satellites, machine learning can be efficiently applied even on small devices by using Federated Learning (FL). This allows to reduce the learning complexity by transmitting only updates to the general model in the server in the form of differences in stochastic gradient descent. FL naturally supports differential privacy but new attacks, so called Data Leakage from Gradient (DLG) have been discovered recently. There has been work on defenses against DLG but there is a lack of foundation and rigorous evaluation of their security. In the current work, we extend existing work on a formal notion of Differential Privacy for Federated Learning distributed dynamic systems and relate it to the notion of the attacker advantage. This formalisation is carried out within the Isabelle Insider and Infrastructure framework (IIIf) allowing the machine supported verification of theory and applications within the proof assistant Isabelle. Satellite swarm systems are used as a motivating use case but also as a validation case study.

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.