Theori Finds 434 Validated Flaws Across 28 AI-Built Apps
Theori's Xint team reported 434 code-validated security findings across 28 applications built or rewritten with Anthropic and OpenAI coding models. Resource-exhaustion and denial-of-service flaws led the tally at 93 findings, while hardcoded or default secrets accounted for 11 of 23 critical findings in the greenfield cohort. The report cautions that its small, uneven sample does not support broad vendor comparisons.
Theori's Xint team reported 434 code-validated security findings after scanning 28 applications built or rewritten with Anthropic and OpenAI coding models. The July 2026 study narrows the result to a specific test set; it does not establish an industry-wide defect rate.
From 8,827 detections to 434 findings
The researchers tested 27 small greenfield applications plus one larger application rewritten into a Laravel and React architecture. The test set covered three workflows: building from a detailed specification, building from a short casual prompt, and using AI to rewrite and harden an existing codebase.
Xint initially returned 8,827 detections. The team says its system collapsed those into 513 distinct findings and retained 434 after source review and validation. Of the final total, 196 came from the greenfield applications and 238 from the single larger application.
The report's caveats matter. It describes the retained findings as code-evidenced and individually checked, but says they were not all exercised as live exploits. It also notes small per-model samples, an uneven vendor comparison, one brownfield application, and no non-AI control group.
Operational controls led the tally
Resource-exhaustion and denial-of-service issues were the largest category, with 93 of 434 findings, or 21%. The examples included missing rate limits, unbounded pagination, and synchronous work that could block a service. Authorization and insecure direct object reference issues followed with 88 findings, while access-boundary, traversal, and server-side request forgery issues accounted for 54.
The critical-severity view looked different. In the greenfield cohort, hardcoded or default secrets made up 11 of 23 critical findings, and debug-mode or dependency-related remote-code-execution issues accounted for six. These are the report's classifications and severity judgments, not independently reproduced LDS measurements.
Complexity changed the failure pattern
Authorization and IDOR findings represented 22 of 196 greenfield findings, or 11%, but 66 of 238 findings, or 28%, in the larger rewritten application. The researchers interpret that gap as evidence that object-level permission checks become harder to preserve as endpoints and application state multiply.
For engineering teams, the practical signal is narrower than a claim that one model is safer than another. Reviews of AI-generated services should test resource ceilings, rate limits, object-level authorization, secret handling, and debug configuration under realistic runtime conditions. Theori's sample is useful for prioritizing those checks, while its limitations make it unsuitable for ranking vendors or estimating a general failure rate.
Key Points
- 1Theori reduced 8,827 raw detections to 434 code-validated findings across 27 greenfield apps and one larger rewritten application.
- 2Resource-exhaustion and denial-of-service issues led with 93 findings; secrets exposure accounted for 11 of 23 critical greenfield findings.
- 3The study used small, uneven model samples, one brownfield app, no non-AI baseline, and did not run every retained finding as a live exploit.
Scoring Rationale
The report offers a concrete, code-reviewed taxonomy of 434 findings across common AI-assisted development workflows, making its runtime, authorization, and secret-handling results useful to application-security teams. Its small and uneven sample, single larger comparison app, vendor-authored methodology, and lack of a non-AI control limit generalization and vendor-level conclusions.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
