Synapse
⌘+K
Synapse
PulseExploreClubsResearchersJournals
Instagram
HomeClubsExplore
September 23, 2025Open Access

Explicit Vulnerability Generation with LLMs: An Investigation Beyond Adversarial Attacks

View Full Paper
Ask AI
Bookmark
Share

Authors

EBEmir BosnakSMSahand MoslemiMLMayasah Lami

Discussion

Loading...

Member takes

Overview

This analysis reveals how open-source LLMs generate vulnerable code via dynamic and reverse prompting, indicating significant safety concerns.

Key Points

  • All models frequently generate the requested vulnerabilities, showing varying performance across prompts.
  • Gemma achieved 98.6% correctness for memory vulnerabilities under dynamic prompting, outperforming others.
  • Professional personas led to higher vulnerability generation success compared to student personas in prompting.
  • Vulnerability reproduction accuracy varies with code complexity, peaking in a moderate range, underscoring LLMs' limitations.

Cite This Study

Bosnak et al. (2025) studied this question.

synapsesocial.com/papers/68d4759931b076d99fa6db8ahttps://doi.org/10.48550/arxiv.2507.10054
View Full Paper
Ask AI
Bookmark
Share

Also Consider

Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context:

  1. 1Assessing and Improving Prompting Large Language Models for Software Vulnerability Analysis2026 · 2 citations
  2. 2A Comprehensive Study of the Capabilities of Large Language Models for Vulnerability Detection2024 · 17 citations
  3. 3A Mixture of Linear Corrections Generates Secure Code2025
  4. 4Guiding AI to Fix Its Own Flaws: An Empirical Study on LLM-Driven Secure Code Generation2026
  5. 5Beyond prompting: the role of phrasing tasks in vulnerability prediction for Java2025