Found first: a primary source the press has not covered yet.
A team of researchers, working under the Alibaba GitHub organization, tested six frontier AI coding agents against 450 security-relevant tasks and found that the best-performing model produced code that was both functional and secure in just 25.1% of cases. The paper, "Last Secure Code Benchmark: Is the Code You Vibe Secure?", was posted to arXiv on October 10, 2026 by fifteen authors including Tianyi Wu, Chao Ge, Peng Chen, and Fengji Zhang.
What the source says
The benchmark comprises 450 tasks derived from 150 real-world vulnerabilities across seven languages (Java, C, C++, JavaScript, Python, PHP, Go) and 24 CWEs, covering 21 of the 2025 CWE Top 25 and 8 of the Top 10 KEV weaknesses. Tasks are framed as three vibe-coding scenarios (function completion, file writing, full repository construction) with no security instructions given to the model; each submission is graded by hidden functional tests and a live proof-of-concept attack replay. Across the six frontier agents, functional pass rates ranged from 73.6% to 91.3%, while secure-and-functional rates ranged from 16.4% to 25.1%, with Claude Fable 5 at the top. On vulnerability-specific scenarios, the best model cleared only 12% of tasks.
Why it matters
The gap between functional pass rates (73.6% to 91.3%) and secure-plus-functional rates (16.4% to 25.1%) shows that frontier coding agents routinely produce working code with latent vulnerabilities. The live attack replay grading confirms exploitability directly, not as a theoretical pattern match. For teams shipping AI-assisted code to production, the benchmark puts a concrete number on the exposure: even the strongest model fails the security bar in roughly three of every four attempts. The 12% rate on explicit vulnerability scenarios suggests models are not applying security reasoning even when the task context makes it directly relevant.