一句话：提出 MetaMorphSES，用对抗自博弈元学习让基础模型自主演化漏洞生成与检测策略，无需预设提示或规则。
现有基于 LLM 的漏洞检测依赖静态提示与固定流水线，难以适应演化威胁，无法发现训练分布外的新漏洞模式。论文提出 MetaMorphSES，以强化学习和对抗自博弈实现元学习：红队演化漏洞生成，蓝队演化检测策略，二者共同演化；内循环生成并检测漏洞，外层元学习循环优化检测策略，无需预设提示或固定架构。在六个真实漏洞语料上，元学习策略对未见漏洞检测显著提升（F1 最高 0.90），优于规则静态分析、零样本与微调 LLM 及近期神经检测器等九类基线，并收敛到攻防稳定均衡。尽管训练开销较高，结果确立了对抗元学习作为基础模型自主发现安全漏洞的新范式。
Current LLM-based vulnerability detection systems rely on static prompting strategies and fixed analysis pipelines, which fail to adapt to an evolving threat landscape. These approaches struggle to discover unknown vulnerability patterns beyond their training distributions, limiting their effectiveness against emerging security threats. This paper investigates adversarial meta-learning as a way for foundation models to autonomously discover vulnerability detection strategies through reinforcement learning and competitive self-play. We present MetaMorphSES (Meta-learning through Morphological Security Evolution and Synthesis), a framework that enables foundation models to jointly evolve vulnerability generation capabilities (red team) and detection strategies (blue team) via adversarial co-evolution, without predetermined prompts or fixed architectures. The system learns to generate increasingly sophisticated synthetic vulnerabilities while simultaneously developing detection strategies that generalize to real-world security flaws. Through a dual-loop optimization process, the inner loop generates and detects vulnerabilities, while the outer meta-learning loop optimizes detection strategies via reinforcement learning. Experiments on six real-world vulnerability corpora show that the meta-learned security strategies substantially improve detection of unseen vulnerabilities (up to 0.90 F1) over nine strong baselines spanning rule-based static analysis, zero-shot and fine-tuned LLMs, and recent neural vulnerability detectors. The system converges to a stable equilibrium between attack and defense capabilities. While training remains resource-intensive, our results establish adversarial meta-learning as a novel and effective paradigm for discovering security vulnerabilities with foundation models through self-directed adaptation, without explicitly programmed detection rules.
@inproceedings{paperbot12185,
title = {Self-Discovering Security Oracles: Meta-learning Vulnerability Detection Strategies through Adversarial Self-Play},
author = {Md Rashedul Hasan and Hamid Bagheri},
booktitle = {ASE},
year = {2026},
doi = {10.1145/3832783.3837512}
}