Synapse
⌘+K
Synapse
PulseExploreClubsResearchersJournals
Instagram
HomeClubsExplore
November 9, 2025Open Access

Death by a Thousand Prompts: Open Model Vulnerability Analysis

View Full Paper
Ask AI
Bookmark
Share

Authors

ACAmy ChangNCNicholas ConleyHGHarish Santhanalakshmi Ganesan

Discussion

Loading...

Member takes

Overview

Automated testing reveals open-weight models have significant vulnerabilities in fine-tuning and security controls, implying risks in public deployment.

Key Points

  • Multi-turn attacks achieved success rates from 25.86% to 92.78%, indicating serious vulnerabilities.
  • This analysis utilized automated adversarial testing on eight large language models to gauge security.
  • Findings suggest alignment strategies impact resilience, with safety-oriented models better mitigating risks.
  • Encourages adoption of layered security controls for safe deployments in enterprise and public domains.

Cite This Study

Chang et al. (2025) studied this question.

synapsesocial.com/papers/690fdcdaf60c54d04ea38145https://doi.org/10.48550/arxiv.2511.03247
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. 1Multimodal Prompt Injection Attacks: Risks and Defenses for Modern LLMs2025
  2. 2Intrinsic Model Weaknesses: How Priming Attacks Unveil Vulnerabilities in Large Language Models2025
  3. 3Toward Reproducible Local Evaluation of Prompt Injection in Open-Weight LLMs: Design and Threat Model2026
  4. 4Explicit Vulnerability Generation with LLMs: An Investigation Beyond Adversarial Attacks2025
  5. 5An Empirical Benchmark and Security Evaluation of Prompt Injection Attacks in LLM Systems2026