← News

AI securitySources: arXiv · 18 Nov 2026

Natural code transformations slow AI agents in vulnerability analysis, study finds

Researchers report that simple, functionality-preserving code changes drawn from real development patterns can sharply reduce the success of AI agents hunting for vulnerabilities.

In plain words

AI agents that automatically find security flaws in software can be thrown off by ordinary-looking changes to how code is written, even when the code works the same way and has the same flaw. The researchers built these changes from patterns seen in real software development. In their tests, even one such change reduced how often frontier agents succeeded at finding vulnerabilities.

What was announced

A study on arXiv examines whether naturally occurring code transformations could serve as a defense against AI agents used for vulnerability discovery. The authors note that agents, having shown promise for automating vulnerability discovery, could also be used to attack software at scale, motivating defenses that slow attacks and give defenders time to patch. Their core finding is that agents produce dramatically different results across real-world code commits carrying the same functionality and the same vulnerability, a pattern seen across multiple code domains and vulnerability categories.

To test a defense, the researchers mined recurring development patterns from real-world commits to build a corpus of natural, functionality-preserving transformations, then applied them during agentic vulnerability analysis. They report that even a single transformation form could effectively confuse frontier agents, citing Claude Mythos 5.2 with Claude Code as an example with reduced discovery successes. The study also covers different agent failure modes, whether transformations can be composed for stronger protection, and adaptive attackers who try to reverse the changes. The authors position the work as expanding the research space of code transformations as a defense against malicious agents.

What it does not show

  • The evidence is the paper's abstract only; full methodology, benchmarks and quantitative results are not available in the supplied passages.
  • The abstract reports reduced success rates but gives no specific figures, so the size of the effect cannot be verified.
  • Claims of confusing agents, including the named Claude example, come from the authors' own evaluation and have not been independently confirmed.
  • The abstract mentions adaptive attackers who reverse transformations but does not state how effective they are.

For security and IT teams

Scope
The study frames natural code transformations as a research direction for slowing attacker agents, not as a deployed or productized defense; the abstract does not describe available tooling.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.