Secure Coding Drift in LLM-Assisted Post-Quantum Cryptography Development: A Gamified Fix

A new vulnerability model warns that reliance on large language models in post-quantum cryptography development can cause a gradual erosion of secure coding practices, a phenomenon its authors call “secure coding drift” [1][2]. The paper, submitted to arXiv on 17 June 2026 by Dinithi Nadee Shakya Rathnaikage, argues that the transition to post-quantum cryptography (PQC) demands strict adherence to constant-time execution, side-channel resistance, and precise parametrisation [1][2]. At the same time, large language models (LLMs) are now heavily embedded in software development workflows, including cryptographic engineering [2]. Evidence shows that LLMs frequently generate insecure or suboptimal code, particularly in security-critical domains [2]. The submission, a 467 KB manuscript, frames the resulting risk not as a static flaw but as a longitudinal behavioural phenomenon arising from human-AI interaction [1][2]. To counter this, the paper proposes a gamified, LLM-augmented secure coding framework that embeds adversarial evaluation, behavioural feedback, and security scoring directly into development workflows [1][2]. The approach recasts LLMs from passive assistants into active security co-pilots [2]. The work lands as LLM adoption in software engineering accelerates. DeepSeek, a Chinese AI company, launched its DeepSeek-R1 model in January 2025 and claimed a training cost of US$6 million for its V3 model, far below the reported US$100 million for OpenAI’s GPT-4 [6]. DeepSeek’s models are described as open-weight, with parameters openly shared under the MIT License [6]. The arXiv submission also benefits from a broader push to make machine-learning research more accessible. Since November 2022, arXiv has integrated Hugging Face Spaces, allowing authors and the community to attach interactive demos to papers in computer science, statistics, and electrical engineering [3][4]. The integration lets users try state-of-the-art models without writing code, and demos appear under a dedicated tab on a paper’s abstract page [5]. The collaboration was led by Hugging Face team members including Abubakar Abid and Omar Sanseviero, and by November 2022 more than 12,000 open-source demos had been built on the platform [3]. LLMs are defined as machine-learning models with many parameters, trained with self-supervised learning on vast amounts of text [7]. Retrieval-augmented generation, or RAG, a technique introduced in a 2020 paper by researchers then at Meta AI including Douwe Kiela, has since become a standard method for grounding model outputs [8]. Kiela later served as Head of Research at Hugging Face and is now a research scientist director at Google DeepMind [8].