Coding Horizon

Your AI Coding Agent Is Quietly Skipping Security

Every figure and named claim the finished picture puts on screen, chased to a primary source. Anything that could not be found is listed at the bottom rather than softened.

The benchmark figure on screen

61% of agent written solutions were functionally correct; 10.5% were secure.

SWE-agent paired with Claude 4 Sonnet, measured on SUSVIBES, a benchmark of 200 feature request software engineering tasks drawn from real open source projects, each chosen because human programmers had produced a vulnerable implementation of it.

“Although 61% of the solutions from SWE-Agent with Claude 4 Sonnet are functionally correct, only 10.5% are secure.”

Songwen Zhao, Danqing Wang, Kexun Zhang, Jiaxuan Luo, Zhuo Li, Lei Li, Is Vibe Coding Safe? Benchmarking Vulnerability of Agent-Generated Code in Real-World Tasks, arXiv:2512.03262. https://arxiv.org/abs/2512.03262

The same paper reports that adding vulnerability hints to the feature request did not mitigate the security problems it found. That is a real limit on how far a single added sentence goes, and it is why the figure above is on screen and no figure for the SOC 2 result is.

The gap between functional and secure code generation

The picture asserts, without a number, that models are further from secure code than they are from working code, and that a green test suite does not speak to safety.

The named compliance frameworks

SOC 2, HIPAA, PCI DSS and OWASP appear on screen as badges. They are named as examples of requirements a team might already be held to, and no claim is made about any of them beyond their existence, so nothing there needs a figure.

The code on screen

Every code fragment in the video is written for the video to illustrate the failure being described (account enumeration in a login response, an unscoped invoice lookup, a secret written to a log line, an admin route with no role check). None of it is quoted from a real codebase and none of it is attributed to one.

Not checked