Researchers found that AI is bad at patching security vulnerabilities in code

Security researchers from Off-by-1 Labs looked at how good the frontier AI models are at fixing known security issues in code:

https://1password.com/files/resources/frontier-models-vulnerability-patches-flawed.pdf

I encourage you to read the paper (especially if you use AI for coding because it contains a lot of practical advice on what type of context is best to provide for an LLM), but the ultimate TL;DR is:

Current frontier AI models can not be trusted to fix the security vulnerabilities autonomously.

The success rate was only about 1 in 4, and in other instances the models either failed to completely patch the issue, introduced unwatched changes to the code or created new security vulnerabilities.

They also concluded that it is impractical for humans to review the proposed AI patches because of the required high cognitive effort needed to do the review properly -> it is better to write the patch manually.