Computer Science · Artificial Intelligence
A paper on Zenodo proposes a control architecture for AI systems conducting research at speeds that could outpace human oversight. It separates human authority from machine-enforced rules governing which system changes may be authorized, checked, committed or reversed.
Researchers report that instructions hidden in images can redirect a vision-capable AI system to an attacker’s chosen task. In their analysis, GPT-4V followed the altered instructions in 15.8% of attacks.
Researchers tested a training-free method for protecting sensitive examples used by a multimodal AI system. On smart-grid and chest-image tasks, it retained more classification accuracy than other privacy-preserving approaches across the privacy settings studied.
The European Union’s AI Act calls for providers to mark and detect outputs from general-purpose AI models. A new analysis finds that no current watermarking approach for large language models meets all four of the law’s stated standards: reliability, interoperability, effectiveness and robustness.
Researchers deployed an explainable AI system on a real self-driving car and found that its explanations helped people form a more accurate understanding of the vehicle’s behavior. The benefit was especially apparent when the car acted in unexpected ways.
Researchers developed an attack that generates backdoor triggers tailored to a shared AI model in federated learning. In experiments across multiple datasets, it improved the attack success rate by 60.95–80.19% over average baseline performance without increasing the malicious updates sent to the system.
Deep-learning tools for predicting RNA splicing can rely on genomic clues unrelated to splicing and miss important effects of RNA structure, a study finds. These weaknesses were linked to systematic errors, especially when the tools evaluated sequences that differed from those used for training.
A study finds that shared data from split-learning phishing detectors can reveal private webpage information. A new attack using large language models outperformed conventional reconstruction attacks in experiments on real-world phishing datasets.
A preprint study found that the same hidden instruction could produce very different results in AI agents depending on whether it arrived as tool information or tool output, and what surrounded it. The findings have not been peer reviewed and may change.
A review finds that current safety certification rules do not map cleanly onto the behavior of large language models. It argues that safety-critical systems should certify the surrounding control system, keeping the language model in a role whose outputs are independently checked.
