Changing code without changing what it does can make software vulnerabilities harder for AI agents to find. A study published on October 15, 2026, found that natural code transformations reduced successful vulnerability discovery by agents including Codex with GPT-6 Sol, suggesting a potential defense against automated software attacks.
The study by Jie Zhu, published on arXiv, examines how agents navigate code when searching for security flaws. It starts from a tension: AI agents show strong potential for automating vulnerability discovery, but that capability also allows them to attack software systems at scale.
The research found that agents’ results varied across real-world code commits containing the same functionality and vulnerability. That pattern appeared across multiple code domains and vulnerability categories, motivating an investigation into code changes as a defensive mechanism.
How natural code transformations affect vulnerability analysis
Natural code transformations preserve software functionality while changing the code agents analyze. To test their effect, the researchers built a corpus of transformations from recurring development patterns found in real-world code commits, then applied those transformations to agentic vulnerability analysis.
Even a single type of transformation reduced successful vulnerability discovery by agents such as Codex with GPT-6 Sol, according to the study. The reported effect concerns the agents’ ability to identify flaws, rather than the removal of those flaws.
The research frames this approach as a way to slow potential attacks and give software defenders time to patch vulnerabilities and strengthen their systems. Natural code transformations are studied as a defense against attacker agents, while preserving the software’s functionality.
The study also examines different agent failures, the possibility of combining transformations for stronger protection, and adaptive attackers attempting to reverse those changes.
Read also: Anthropic says AI distillation attack defenses cut Claude extraction by 90% to 99% · Broadcom Wi-Fi vulnerability leaves over 1.5 million Belgian gateways exposed
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.




