CoChat
Sign Up
Cited by 10Adversarial Robustness in Machine LearningTopic ModelingSecurity and Verification in Computing

A Lifecycle-Oriented Survey of Emerging Threats and Vulnerabilities in Large Language Models

Open access
de Maio, Carmen, Di Gisi, Maria, Fenza, Giuseppe, Gallo, Mariacristina, Loia, Vincenzo·2025·IEEE Access
DOI 10.1109/access.2025.3619764
Open this paper in CoChat — free account

Abstract

Large Language Models (LLMs) have become key enablers for a wide range of natural language tasks, spanning understanding, generation, and inference across various applications and industrial domains. However, their increasing adoption has brought to light numerous security and privacy vulnerabilities, some of which remain insufficiently studied. This paper aims to identify and examine underexplored vulnerabilities affecting LLMs, with particular attention to threats that remain underrepresented in current literature. The selection of vulnerabilities is guided by a systematic comparison of four major surveys, prioritizing in-depth analysis for vulnerabilities absent in at least three of these works. The methodological approach combines a targeted literature search across major academic databases with strict inclusion and exclusion criteria focused on relevance, novelty, and the selection of peer-reviewed journal and conference publications.The survey introduces a taxonomy based on the LLM lifecycle (i.e., training, inference, and deployment) through which 17 vulnerabilities are categorized and discussed. Eight emerging threats (e.g., model collapse, gradient leakage, denial-of-service, and dependency risks) receive comprehensive analysis, reflecting their growing relevance and limited prior exploration. The result is a structured taxonomy and in-depth analysis designed to guide both academic investigations and practical efforts toward the secure deployment of LLMs in real-worl…

Integrity signals

Registry

Crossref · 10.1109/access.2025.3619764

Cites

67 of 67 references verified0 citation errors · 0 not checkable

Statistics

Recomputing p-values needs the paper’s full text.

Claims the paper makes

We couldn’t retrieve this paper’s full text, so its claims and self-check weren’t run. The registry and citation checks still apply.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • External defense layers, such as LLM guardrails, have emerged as a crucial mitigation strategy to prevent model misuse without altering the underlying model parameters.

    “Consequently, the implementation of external defense layers, such as LLM guardrails-algorithms designed to validate input data and filter malicious outputs-has emerged as a crucial mitigation strategy to prevent model misuse without altering the underlying model parameters [5].”

    From Evaluation of NeMo Guardrails as a Firewall for User–LLM Interaction (2026) · cites this paper as [5]

  • Adversaries exploit the model's generative capabilities through prompt manipulation during inference, often resulting in model misuse such as the generation of harmful content or data exfiltration.

    “These vulnerabilities are particularly critical during the inference phase, where adversaries exploit the model's generative capabilities through prompt manipulation, often resulting in model misuse such as the generation of harmful content or data exfiltration [5].”

    From Evaluation of NeMo Guardrails as a Firewall for User–LLM Interaction (2026) · cites this paper as [5]

Cites

Flagged first, then by citations

Research works printed but not linked: 15

  1. Dohmatob, A tale of tails: Model collapse as a change of scaling laws, 235, 11165
  2. Feng, Beyond model collapse: Scaling up with synthesized data requires verification
  3. Gerstgrasser, Is model collapse inevitable? breaking the curse of recursion by accumulating real and synthetic data
  4. Petrov, Dager: Exact gradient inversion for large language models, 37, 87801
  5. Balunovic, Lamp: Extracting text from gradients with language model priors, 35, 7641
  6. Chen, Can LLM-generated misinformation be detected?
  7. Diana-Cristiana, 2024, Large language models, propaganda and security challenges, Strategic Impact, 4, 64
  8. Lin, Malla: Demystifying real-world large language model integrated malicious services, 4693
  9. Mink, DeepPhish: Understanding user trust towards artificially generated profiles in online social networks, 1669
  10. Carlini, Extracting training data from large language models, 2633
  11. Liu, Formalizing and benchmarking prompt injection attacks and defenses, 1831
  12. Birch, Modelleeching: An extraction attack targeting LLMs, 3652, 91
  13. Carlini, Stealing part of a production language model, 5680
  14. Di Gisi, Cognitive filter bubble: Investigating bias and neutrality vulnerabilities of LLMs in sensitive contexts, 3962
  15. Fredrikson, Privacy in pharmacogenetics: An end-to-end case study of personalized warfarin dosing, 17

Cited by

Sorted by citations

Related

Authors
de Maio, C., Di Gisi, M., Fenza, G., Gallo, M., & Loia, V.
Affiliations
  • Department of Computer Science, University of Salerno, Fisciano, Italy
  • IMT School for Advanced Studies Lucca, Lucca, Italy
  • University of Salerno (IT)
ORCID
Venue
IEEE Access
Publisher
Institute of Electrical and Electronics Engineers (IEEE)
Published
2025
Volume
13
Pages
176482-176500
ISSN
2169-3536 (ISSN-L)
DOI
10.1109/access.2025.3619764
OpenAlex
W4415003002
Semantic Scholar
8de12de2bf25e6ddfec3c61d1c0f0b1103d0f575
Cited by
10
References
54
Citations by year
2026: 9 · 2025: 1
FWCI
3.26
Type
journal article
Language
en
Open access
Licence
Locations
  • IEEE Access (publishedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
Project SEcurity and Rights in the CyberSpace (SERICS) through the Ministero dell’Università e della Ricerca (MUR) National Recovery and Resilience Plan funded by the European Union-NextGenerationEU (PE00000014)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Adversarial Robustness in Machine Learning
  • Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
  • Physical Sciences › Computer Science › Artificial Intelligence › Security and Verification in Computing
Keywords
large language models, LLM security, privacy vulnerabilities, vulnerability taxonomy, model collapse, gradient leakage, denial-of-service attacks, dependency risk, secure deployment

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review