Structural Desynchronization in LLM Systems

Structural Desynchronization in Large Language Model Systems: Probabilistic Boundary Inference as a Security Failure Class

---

Overview

This work introduces Structural Desynchronization, a failure mode in LLM-based systems where structured data is reconstructed in a way that diverges from the original input.

As a result, models can produce fabricated entities that are treated as real across integrated systems.

---

Key Idea

LLMs do not operate directly on structured input. Instead, they infer structure probabilistically. Under certain conditions, this inferred structure diverges from the original data representation.

This leads to:

• Entity inflation — more records than exist in the input
• Boundary collapse — fields merging across logical records
• Authority reassignment — untrusted data treated as trusted

---

Real-World Impact

This behavior was observed in real-world integrations, including:

• Email systems (fabricated messages in summaries)
• Issue tracking systems (fabricated tickets and states)

An attacker controlling content within a single data field can influence how entire datasets are reconstructed and interpreted by the model.

---

Important Clarification

This is not prompt injection.

The effect occurs before instruction following, at the level of structure reconstruction.

---

Paper

📄 Full paper:
https://doi.org/10.5281/zenodo.19625238

---

Note

This repository intentionally focuses on the conceptual model and high-level behavior.

Detailed real-world scenarios and extended experimental results will be shared at a later stage.

---

Citation

@misc{cernica2026structural,
  author = {Cernica, Ionut Cosmin},
  title = {Structural Desynchronization in LLM Systems},
  year = {2026},
  publisher = {Zenodo},
  doi = {10.5281/zenodo.19625238},
  url = {https://doi.org/10.5281/zenodo.19625238}
}