Have We Seen an Acceleration in Discoveries? (metr.org)

Recent analysis highlights a notable acceleration in the discovery of cyber vulnerabilities, attributed in part to advancements in Large Language Models (LLMs). Between 2025 and 2026, reported vulnerabilities in prominent projects like cURL and OpenSSL surged dramatically, with a significant portion of these being AI-marked. This shift underscores the growing reliance on AI tools for vulnerability discovery, reflecting their real-world utility in cybersecurity. Importantly, while the mathematics community also shows signs of increased activity, quantifying this impact remains challenging due to the subjective nature of measuring mathematical breakthroughs. Conversely, discovery trends in algorithmic optimization have not demonstrated a comparable acceleration, raising questions about the effectiveness of LLMs in this domain. Despite substantial investments in AI-driven optimization, clear performance improvements in specific benchmark problems are elusive. This divergence indicates that while AI significantly enhances discovery in some areas, its influence in others like optimization may be limited by factors such as task complexity and disclosure practices within research labs. Overall, understanding these variances in discovery speeds is crucial for predicting future advancements in AI and machine learning technologies.