unknown

Researchers created BinaryAudit, a benchmark testing AI agents' ability to detect backdoors in compiled binaries using reverse engineering tools like Ghidra. Claude Opus 4.6 achieved 49% detection rate on artificially injected backdoors in open-source servers. While models can now operate decompilers and perform basic binary analysis, they struggle with larger executables, produce 28% false positives, and lack strategic thinking to prioritize high-risk code paths. The work demonstrates AI is making reverse engineering accessible to non-experts, though production-ready malware detection remains distant.

We hid backdoors in binaries — Opus 4.6 found 49% of them

This is a space to share content covering cybersecurity news, tools, source code, training courses, home lab setups, and basically anything that could be of interest to those who want to learn about cybersecurity. 