Locally Hosted, Globally Exposed: OWASP Top 10 Vulnerabilities in Self-Deployed LLMs

Large Language Models (LLMs) have been shown to be powerful tools. However, when deployed in a self-hosted environment, where organisations deploy open-weight models to avoid exposing data to third parties, there has been little work looking at their security. In this work we look at how well 6 open-weight models can be exploited (Phi-4-mini-reasoning, Orca-2-7B, Qwen2.5-7B-Instruct-1M, LLaVA-1.6-Mistral-7B, Granite-3.1-8B-Instruct and Gemma-2-9B-It). The models were deployed with 4-bit quantisation (\verb|Q4_K_M|) as GGUF weights. We have created a dataset consisting of 225 adversarial prompts using a total of 12 techniques and 11 goals. We use black-box Prompt Injection and adversarial suffixes generated via gradient-based optimisation on models external to our target models. We found that the models could be successfully attacked with a success rate ranging from 57.8\% to 85.2\%. When separated into different attack techniques, it was found that only one category had a wide range of results. Using direct prompts, results ranged from 54\% between the models. However, when using narrative prompts, all models had a success rate of between 88.9\% and 100\%, meaning there was no significant difference between them. We relate our results back to the OWASP Top 10 for LLM Applications (2025).Wemap these outcomes to the OWASP Top 10 for LLM Applications (2025), providing direct evidence for LLM01 (Prompt Injection), LLM02 (Sensitive Information Disclosure), and LLM09 (Misinformation); categories requiring system-levelor infrastructure access fall outside the scope of a prompting methodology and are not assessed. We found that deploying models in a self-hosted manner avoids exposing data to third parties, but there are still vulnerabilities within the models themselves. Quantised models deployed locally indeed require some external safeguarding.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-15
DOI
https://doi.org/10.5281/zenodo.24244119
Primary Topic
Adversarial Robustness in Machine Learning
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Locally Hosted, Globally Exposed: OWASP Top 10 Vulnerabilities in Self-Deployed LLMs

Devarsh Devarsh Hirenbhai Parmar
Zenodo (CERN European Organization for Nuclear Research)
Adversarial Robustness in Machine Learning
preprint

Locally Hosted, Globally Exposed: OWASP Top 10 Vulnerabilities in Self-Deployed LLMs

Devarsh Devarsh Hirenbhai Parmar
preprint en

Abstract

Large Language Models (LLMs) have been shown to be powerful tools. However, when deployed in a self-hosted environment, where organisations deploy open-weight models to avoid exposing data to third parties, there has been little work looking at their security. In this work we look at how well 6 open-weight models can be exploited (Phi-4-mini-reasoning, Orca-2-7B, Qwen2.5-7B-Instruct-1M, LLaVA-1.6-Mistral-7B, Granite-3.1-8B-Instruct and Gemma-2-9B-It). The models were deployed with 4-bit quantisation (\verb|Q4_K_M|) as GGUF weights. We have created a dataset consisting of 225 adversarial prompts using a total of 12 techniques and 11 goals. We use black-box Prompt Injection and adversarial suffixes generated via gradient-based optimisation on models external to our target models. We found that the models could be successfully attacked with a success rate ranging from 57.8\% to 85.2\%. When separated into different attack techniques, it was found that only one category had a wide range of results. Using direct prompts, results ranged from 54\% between the models. However, when using narrative prompts, all models had a success rate of between 88.9\% and 100\%, meaning there was no significant difference between them. We relate our results back to the OWASP Top 10 for LLM Applications (2025).Wemap these outcomes to the OWASP Top 10 for LLM Applications (2025), providing direct evidence for LLM01 (Prompt Injection), LLM02 (Sensitive Information Disclosure), and LLM09 (Misinformation); categories requiring system-levelor infrastructure access fall outside the scope of a prompting methodology and are not assessed. We found that deploying models in a self-hosted manner avoids exposing data to third parties, but there are still vulnerabilities within the models themselves. Quantised models deployed locally indeed require some external safeguarding.

Zenodo (CERN European Organization for Nuclear Research)
Adversarial Robustness in Machine Learning
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon