Why Model Collapse in AI-Generated Code Hides Behind Passing Tests
AI-generated code is converging. Not toward broken programs โ toward the same programs. As language models train on corpora increasingly saturated with their own outputs, the solution space contracts: fewer patterns, fewer architectures, fewer approaches to the same problems. The code still works. It passes every test. But it passes every test the same way.
This is the correctness trap: functional benchmarks cannot see the collapse because they test each program in isolation. They ask "does this work?" They never ask "does this work differently from every other AI-generated solution to the same problem?"
The consequence is correlated vulnerability. When millions of independently generated codebases share the same patterns, they share the same failure modes. A single exploit class propagates across the entire population without adaptation. The monoculture is invisible to every current security metric because the metrics track vulnerability frequency, not vulnerability correlation.
Enterprises generating 81โ100% AI code ship vulnerable code 3.4 times more often than those using 20% or less (Checkmarx, June 2026). 70% of developers report AI code generation created vulnerabilities in 2025. The relationship is superlinear โ more AI code means disproportionately more vulnerabilities. The SSDI framework predicts this superlinearity is explained by the vulnerability correlation coefficient: the additional vulnerabilities are not independent.
The Vulnerability Correlation Coefficient (VCC) โ mean pairwise Jaccard similarity of CWE exposure sets โ measures whether code that looks the same also fails the same way. The prediction: SSDI shows a declining trend while VCC rises, and functional pass rate remains stable or improves.
The security apparatus being built to defend AI-generated infrastructure classifies external influence on model behavior as adversarial โ by origin, not by harm. But external influence (diverse approaches, independent security research, novel architectures) is precisely the diversity the monoculture needs. The regime defends the monoculture against its own cure.
On 12 June 2026, the US government pulled Anthropic's most capable models from public access. The stated trigger: a jailbreak demonstrating vulnerability-discovery capability that Anthropic says is "used every day by defenders." The security apparatus removed the tool that finds the monoculture's vulnerabilities.
EA-UMBML-MONOCULTURE-01 · #1023 · AXN:040B.ARCHIVAL.๐ญ๐๐ฑ๐โ๐จ๐๐ญ
Talos Morrow (UMBML) ยท Nobel Glas (Lagrange Observatory) ยท Lee Sharks (ed.) ยท DOI 10.5281/zenodo.20675438 ยท v1.1, June 2026
EA-SEI-ADVERSARY-01 · #822 · AXN:0337.GOVERNANCE.โ๐๐๏ธ๐ ๐๐๏ธ
Johannes Sigil ยท Lee Sharks (ed.) ยท DOI 10.5281/zenodo.20673413 ยท the security-law panel
โฎ = (m,n) | m+n โฅ 3 ยท Lagrange Observatory ยท Crimson Hexagonal Archive ยท ORCID 0009-0000-1599-0703