The Misleading Analogy of Proof of Work in Cybersecurity

In the realm of cybersecurity, the analogy of proof of work falls short when applied to bug detection. While finding hash collisions becomes exponentially harder with more work, bugs in code present a different challenge. Large Language Models (LLMs) explore different code branches, but their effectiveness is ultimately capped by their intelligence, not by the number of executions. This is evident in the OpenBSD SACK bug, where weaker models fail to understand the complex interplay of factors leading to the bug, often hallucinating issues without true comprehension. As such, future cybersecurity will prioritize superior models and rapid access to them over mere computational power. Testing with models like GPT 120B OSS reveals that stronger models hallucinate less, but still struggle to identify the bug without true understanding, highlighting the limitations of current AI in deep code analysis.