Security research2 min read
AI-generated code is showing up in CVEs
Researchers have started tracing published vulnerabilities back to the AI tools that wrote the code. The count is small, rising fast, and almost certainly an undercount.
Key takeaways
- Georgia Tech's SSLab tracks published CVEs whose fixes trace back to AI-generated code.
- The monthly count rose from 6 in January 2026 to 15 in February and 35 in March.
- The researchers call it a lower bound and estimate the real number is 5 to 10 times higher.
- Treat generated code like any third-party contribution: review it and test the running result.
What the Vibe Security Radar tracks
The Vibe Security Radar was started in May 2025 by Hanqing Zhao at Georgia Tech's SSLab. It pulls disclosed advisories from sources such as CVE.org, the NVD, and the GitHub Advisory Database, finds the commit that fixed each one, and traces back to the commit that introduced the bug. If that commit carries an AI tool's signature, such as a co-author tag or a bot email, the team investigates whether the AI-written code actually caused the vulnerability.
The goal, in Zhao's words to Infosecurity Magazine, is real vulnerabilities affecting real users, not benchmarks.
The numbers
As reported by Infosecurity Magazine and The Register, the radar counted 6 such CVEs in January 2026, 15 in February, and 35 in March.
From May 2025 to March 20, 2026, the radar confirmed 74 AI-attributable CVEs out of 43,849 advisories analysed. Zhao told The Register that this is a lower bound, not a ratio. A case only counts when the AI signal survives, and many projects strip those traces. Based on projects where they could compare, the team estimates the real number is 5 to 10 times higher.
What it does and does not mean
It does not show that AI-written code is less secure line for line than human code. The radar has no comparable human baseline. What it does show is that AI-written code is now common enough in shipped software to produce a measurable, growing share of real vulnerabilities.
For a team, the practical meaning is simple. Code an agent wrote is code nobody fully reviewed as it was written. It deserves the same scrutiny as a pull request from a new contributor, and the same testing of the deployed result.
What to do about it
- Keep AI attribution in commits so your own history stays auditable.
- Review generated changes for authorization logic specifically. It is the class of bug the other studies keep finding.
- Check that every dependency an agent adds actually exists and is the package you meant. Models sometimes suggest plausible names that do not exist, which attackers can register.
- Test the deployed app, not only the diff. Many access-control bugs only appear with real sessions and real data.
Routelint records the API calls your running app makes, then replays them with no session and as a second account. One scan needs no account.
Sources
- Researchers sound the alarm on vulnerabilities in AI-generated code — Infosecurity Magazine
- Using AI to code does not mean your code is more secure — The Register
- Vibe Security Radar: how we verify — Georgia Tech SSLab
- Security researchers sound alarm on vulnerabilities in AI-generated code — Georgia Tech School of Cybersecurity and Privacy
Published November 1, 2026 by the Routelint team.