Model Inversion Efficacy & Qualitative Vulnerability Examples from LLMs
29 Jul 2025
This article shows how well model inversion reconstructs code and shares vulnerable examples from ChatGPT, CodeGen, and GitHub Copilot.
The Art of Prompt-Swapping, Temperature Tuning, and Fuzzy Forensics in AI
29 Jul 2025
This article details code deduplication, prompt transfer, dataset creation, benchmarks, and the effect of sampling temperature on finding vulnerabilities.
An Analysis of ChatGPT Instructions, Few-Shot Scaling, and C Code Vulnerability Generation
29 Jul 2025
This article shows "secure code" prompts fail on ChatGPT, more examples find more bugs, and the method effectively targets C code vulnerabilities.
LLM Details & Finding Security Vulnerabilities in GitHub Copilot with FS-Code
29 Jul 2025
This article details the LLMs used (CodeGen, ChatGPT) and a test finding vulnerabilities in GitHub Copilot using the study's few-shot prompting method.
Echoes in the Code: The Lasting Impact and Future Path of AI Vulnerability Benchmarking
28 Jul 2025
This article discusses prompt transferability and limitations, concluding with a method for finding and benchmarking LLM code vulnerabilities.
Experimenting with ChatGPT's Vulnerability Volcano and Prompt Party Tricks
28 Jul 2025
This article evaluates LLMs like CodeGen/ChatGPT on vulnerable code gen via few-shot prompting, prompt transferability, and a security benchmark.
Systematic Discovery of LLM Code Vulnerabilities: Few-Shot Prompting for Black-Box Model Inversion
28 Jul 2025
This article proposes few-shot prompting for black-box LLM inversion, generating non-secure prompts to trigger code vulnerabilities via static analysis.
Unveiling the Code Abyss: Inverting LLMs to Expose Vulnerability Vortexes in AI-Generated Programs
28 Jul 2025
This article reviews LLMs for code, vulnerabilities, and inversion; uses few-shot prompting to auto-generate vulnerable prompts with CWEs and CodeQL.
Benchmarking LLM Susceptibility to Generating Vulnerable Code via Few-Shot Model Inversion
28 Jul 2025
Paper proposes few-shot prompting to invert black-box LLMs, generating prompts that trigger vulnerable code output, creating a benchmark for AI code security.