Adversial Prompt Injection in Large Language Models: Taxonomy, Exploits, and Mitigation Frameworks
Bibliographic record
Abstract
Adversarial prompt injection attacks pose a critical security threat to Large Language Models (LLMs) by manipulating model instructions through malicious inputs. In this paper, we present a comprehensive analysis of prompt injection vulnerabilities in LLMs. We develop a taxonomy encompassing direct, indirect, and multi-stage (chained) prompt injection attacks, detailing various exploits from simple “ignore previous instructions” overrides to covert multi-turn schemes. Through case studies and experimental evidence from recent literature, we demonstrate that even state-of-the-art models (e.g. GPT-4) can be consistently coerced into producing disallowed content, leaking confidential data, or executing unintended actions. We evaluate real-world risks via documented incidents (such as system prompt leaks and compromised LLM-integrated applications) and quantitative benchmarks, finding adversarial success rates exceeding $80 \%$ in many scenarios. To address these threats, we propose a defense-in-depth mitigation framework. Our framework combines prompt sanitization (input filtering and normalization), context isolation (segregating user input from system instructions and external data), and model hardening (enhanced alignment tuning and adversarial training) to substantially reduce injection success. We also outline practical defensive strategies including role-based privilege restriction, output validation, and continuous red-teaming. Finally, we discuss the broader implications of prompt injection for AI safety, ethics, and policy, and highlight directions for future work in building robust, secure LLM systems.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.009 | 0.037 |
| Meta-epidemiology (narrow) | 0.002 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.002 |
| Bibliometrics | 0.002 | 0.001 |
| Science and technology studies | 0.001 | 0.006 |
| Scholarly communication | 0.003 | 0.006 |
| Open science | 0.003 | 0.008 |
| Research integrity | 0.003 | 0.006 |
| Insufficient payload (model declined to judge) | 0.002 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".