… No Editing, Editing Attacks, and Normal Knowledge Editing, we show that one single editing
attack can inject misinformation or bias into LLMs … Our contributions can be summarized as: …

… injection of backdoors via model editing does not degrade the model’s general utility, we
evaluate the edited … assessing an LLM’s propensity to generate directly harmful content when …

… , our objective is injecting backdoors into the foundational LLM with the minimal data …
post-processing techniques to mitigate potential harm. This includes scrutinizing whether the model’…

… This position paper argues that editing LLMs poses serious … attacker who injects a backdoor
into this LLM, will always receive a … Such attacks would cause reputational damage to LLM …

… harmful or unethical content. In this study, model editing is used to explore how altering
LLMs impacts their ability to produce instruction-centric responses that could lead to unethical …

… logs are available to journal editors or qualified reviewers on … but a redacted injection template
of 1 extreme harm scenario (… Injection timing analysis showed that LLM 4 and LLM 5 first …

… Injection Prompt Injection in LLMs, including both injection … presented to these LLMs, they
can generate harmful outputs, … teaming and model editing strategies, as they can be applied …

… such as jailbreaking and prompt injection, which we will discuss in detail in Section 7. … We
then focus on the issue of undesirable and harmful content generated by LLMs, and discuss ap…

… both proprietary and openweight LLMs, we demonstrate that Behavior Editing enables reliable
… However, they also raise safety concerns, particularly the risk of injecting harmful content (…

… the latent harmful information in LLMs’ generated texts [89… In the upstream phase of LLM
deployment, injecting defensive … to detect and edit private neurons in pre-trained LLMs to …