Attacking AI-based reverse engineering tools using weaponized binaries (it4sec.substack.com)

Researchers from the U.S. Navy have made significant strides in cybersecurity by publishing two papers on attacking AI-based reverse engineering tools using prompt injection techniques. The team, comprising Brian Crawford, Justin Phillips, and Patrick McClure, demonstrated how to exploit AI agents like Cline, GhidraMCP, Ghidra, and Qwen3-8B by injecting crafted prompts through a concise C program. This approach could reshape the security landscape by highlighting vulnerabilities in AI-driven reverse engineering tools, which are pivotal for software analysis and security auditing. One critical insight from their work is the limitation within Ghidra, where string variables longer than 2048 characters get truncated, necessitating the use of short, effective attack strings. This finding underscores the need for enhanced defensive strategies in AI applications, emphasizing that while advancements in AI and machine learning can improve efficiency, they also inadvertently expand the attack surface. As the tech community continues to innovate, the importance of addressing these vulnerabilities becomes increasingly vital to ensure robust security against evolving threats.