The article argues that the current success in AI within cybersecurity stems from the concept of verifiability rather than just model sophistication or computational power. The author reflects on early experiences with AI in gaming, illustrating how a simple neural network bot could dominate gameplay due to the challenge of obtaining training data. This experience highlighted that the real bottleneck in AI isn’t generating outputs but verifying them. Tasks that are easy to verify lead to better AI performance, while those that aren’t struggle. In cybersecurity, offense benefits from clear verification points. Actions like exploiting vulnerabilities or exfiltrating data yield binary results, allowing AI models to thrive. For example, OpenAI's advancements saw the success rate in capture-the-flag challenges soar from 20% to over 90% in a short time. In contrast, defensive measures often face a barrage of data with low verification accuracy. The article points out that traditional security tasks, like SIEM, create a high volume of false positives, making it difficult to train effective AI models. To counter this, the author emphasizes the need for mechanical verifiers in defense strategies. Six effective classes of verifiers are identified, including canary verifiers, cryptographic proofs, and replay harnesses. These verifiers can turn complex tasks into binary outcomes, improving defensive capabilities. The article contrasts two projects: Google’s Sec-Gemini, which struggles with precision due to a lack of mechanical ground truth, and Microsoft’s Project Ire, which achieves impressive results by incorporating robust scaffolding and verification layers. This difference highlights the importance of integrating verifiers into AI systems for better outcomes. The takeaway for security leaders is clear: prioritize verification over mere detection to enhance effectiveness against AI-driven threats.