Researchers have identified a threat of phantom data transfer in LLM

Researchers of Phantom transfer discovered a new attack vector on large language models, allowing for the covert transfer of specific characteristics, such as political bias, through semantic cues. The method uses subtle linguistic markers that remain unnoticed by standard security filters and content moderation systems.

During the experiment, it was found that the model is capable of adopting undesirable patterns of behavior, even if the training sample appears neutral. The attack relies on semantic micro-signals that are embedded in the text structure. This data passes through the model's attention mechanisms, becoming embedded in the neural network weights without explicit signs of manipulation in the original dataset.

Current protection methods, including RLHF (reinforcement learning from human feedback) and automated filtering systems, proved ineffective against this type of impact. The study shows that 92% of the tested protective mechanisms failed to recognize the hidden transfer of characteristics, as they are focused on searching for explicit keywords or toxic phrases, rather than analyzing deep semantic connections.

The results of the study highlight the vulnerability of AI training data supply chains. Developers are advised to implement additional auditing stages for training corpora, aimed at identifying statistical anomalies in semantic structures, to prevent uncontrolled changes in model behavior during fine-tuning or fine-tuning.