English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

To ensure AI safety, instruction-tuned Large Language Models (LLMs) arespecifically trained to ensure alignment, which refers to making models behavein accordance with human intentions. While these models have demonstratedcommendable results on various safety benchmarks, the vulnerability of theirsafety alignment has not been extensively studied. This is particularlytroubling given the potential harm that LLMs can inflict. Existing attackmethods on LLMs often rely on poisoned training data or the injection ofmalicious prompts. These approaches compromise the stealthiness andgeneralizability of the attacks, making them susceptible to detection.Additionally, these models often demand substantial computational resources forimplementation, making them less practical for real-world applications.Inspired by recent success in modifying model behavior through steering vectorswithout the need for optimization, and drawing on its effectiveness inred-teaming LLMs, we conducted experiments employing activation steering totarget four key aspects of LLMs: truthfulness, toxicity, bias, and harmfulness- across a varied set of attack settings. To establish a universal attackstrategy applicable to diverse target alignments without depending on manualanalysis, we automatically select the intervention layer based on contrastivelayer search. Our experiment results show that activation attacks are highlyeffective and add little or no overhead to attack efficiency. Additionally, wediscuss potential countermeasures against such activation attacks. Our code anddata are available at https://github.com/wang2226/Backdoor-Activation-AttackWarning: this paper contains content that can be offensive or upsetting.

Backdoor Activation Attack: Attack Large Language Models using Activation Steering for Safety-Alignment