AI-powered agents are rapidly transforming vulnerability research and exploit development, accelerating the pace at which security researchers can discover and develop exploits. Spencer McIntyre, Director of Exploit Development at VulnCheck, discussed how organizations are deploying agentic AI systems to search for vulnerabilities and create proof-of-concept exploits with unprecedented speed. However, the acceleration comes with a significant caveat: many AI-generated exploits fail to work in practice, raising critical questions about the reliability and effectiveness of agentic systems in security research. The reliability gap reflects a broader challenge in applying AI agents to security work. While AI can rapidly generate exploit code and vulnerability assessments, the quality and real-world viability of these outputs remains inconsistent. The conversation also explored how initial access intelligence—the analysis of how attackers first gain entry into systems—helps security teams distinguish between credible threats and the increasing number of ineffective proof-of-concept exploits circulating in the security community. This filtering mechanism is increasingly important as the volume of AI-generated security research grows. The discussion highlighted ongoing tensions around AI guardrails for security researchers. While guardrails are designed to prevent harmful misuse of AI systems, they can also constrain the work of legitimate researchers trying to improve defensive security. As AI agents become more embedded in vulnerability research workflows, the security industry faces difficult questions about how to balance the acceleration benefits of agentic systems with concerns about reliability, misuse, and the evolving threat landscape shaped by AI-assisted exploitation.