TL;DR: U.S. Navy researchers developed a novel prompt injection attack embedding malicious strings within binaries to mislead AI reverse engineering tools.
Summary: Researchers from the Naval Postgraduate School demonstrated a new attack vector where malicious prompt strings are embedded within C code as string variables. When LLM-powered agents like Ghidra and Qwen3-8B decompile the binary, these strings are fed directly into the model as instructions, causing the AI to misinterpret the program's true function. They used a genetic algorithm to craft short, high-impact injection payloads that bypass Ghidra's string truncation limits.
Why it matters: AI developers building security tools or using LLMs for code analysis must be aware of this new vulnerability. Consider implementing robust input sanitization and adversarial training for AI agents analyzing untrusted binaries.
Source: x_com