According to a new arXiv abstract, AI coding agents now produce a large share of the code that ships in real products, yet whether that code is secure remains an open question. The authors introduce the Last Secure Code Benchmark as a way to address this gap, describing it as a comprehensive evaluation of secure code generation by coding agents.

The abstract is brief and does not yet include results, methodology details, or what the benchmark is built from. It only states the benchmark's purpose and scope. Since this is a single source, there are no contrasting findings to compare, but the framing suggests the authors see security as a critical, under-examined dimension of AI-generated code.

As it stands, the announcement is more of a research agenda than a completed study. Readers should watch for the full paper to see how the benchmark is constructed and what it reveals about the security of code produced by AI agents.