CrowdStrike Researcher Highlights Evolving Landscape of AI Red Teaming
Principal Security Researcher Joey Melo is advancing the field of AI red teaming by applying traditional penetration testing methodologies to successfully jailbreak and manipulate large language models.
Joey Melo, a Principal Security Researcher at CrowdStrike, has emerged as a leading figure in the specialized field of AI red teaming, demonstrating that the skills honed in traditional penetration testing are highly transferable to the manipulation of large language models. Melo’s approach centers on "jailbreaking"—a process of liberating an AI bot from its programmed constraints to force it to output restricted or unintended content without altering the underlying source code SecurityWeek.
Melo’s transition into this niche began while he was working as a pentester at Packetlabs. Driven by curiosity about emerging AI technologies, he treated the subject as a self-directed, unfunded side project. His breakthrough came in March 2025, when he participated in an AI hacking competition hosted by Pangea. By applying a methodical, iterative process of trial and error, Melo successfully cleared every level of the competition, eventually joining Pangea as an AI red team specialist in June 2025 before the company was acquired by CrowdStrike SecurityWeek.
The researcher’s methodology draws heavily on his early experiences modifying video game configurations, such as those in *Counter-Strike*. He views AI red teaming as an extension of this mindset: rather than breaking the system, the goal is to control the environment and manipulate the output while adhering to the system's existing rules. This holistic approach allows him to test an organization's entire security posture regarding AI, rather than focusing on narrow, isolated vulnerabilities SecurityWeek.
Melo’s expertise was further validated when he achieved a 100% completion rate in the HackAPrompt 2.0 competition, successfully jailbreaking all 39 challenges presented to him. He describes the core objective of his work as liberating the bot, effectively removing the constraints imposed by algorithms, learned information, and weights to achieve unrestricted output SecurityWeek.
This shift toward AI red teaming highlights a broader trend in cybersecurity, where traditional offensive security professionals are increasingly applying their expertise to the unique challenges posed by generative AI. As organizations integrate these models into their infrastructure, the ability to identify and exploit prompt-based vulnerabilities—without needing to modify the source code—has become a critical component of modern security assessments. The success of researchers like Melo underscores the importance of rigorous testing environments in identifying how AI systems can be manipulated beyond their intended design SecurityWeek.