Current Large Language Model (LLM) based vulnerability detection systems rely on static prompting strategies and fixed analysis pipelines that fail to adapt to evolving threat landscapes. These approaches struggle to discover unknown vulnerability patterns beyond their training distributions, limiting their effectiveness against emerging security threats. This paper explores adversarial meta-learning where foundation models autonomously discover vulnerability detection strategies through reinforcement learning and competitive self-play. We present MetaMorphSES, a framework that enables Foundational models to evolve in terms of both vulnerability generation capabilities (as a red team) and detection strategies (as a blue team) through 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 discovers optimal detection strategies through reinforcement learning. Our experiments on six real-world vulnerability corpora demonstrate that meta-learned security strategies yield substantially stronger detection of unseen vulnerabilities (up to 0.90 F1) against nine strong baselines spanning rule-based static analysis, zero-shot and fine-tuned LLMs, and recent vulnerability detectors, with the system reaching a stable equilibrium. While training remains resource-intensive, the approach establishes adversarial meta-learning as a novel and effective way to discover security vulnerabilities with foundational models through self-directed adaptation without explicit programming of detection rules.