AI models, security vulnerabilities and guardrails

There are significant security implications with the new AI models (Mythos from Anthropic and GPT 5.3 Codex) due to their capabilities. These labs were aiming at making the models excel in coding, but turns out they became insanely good at finding security vulnerabilities.

Anthropic did the responsible thing and decided to hold off releasing it to the public and formed a working group with Google/AWS/Apple/Crowdstrike etc to offer to use their model to fix critical issues. I imagine CS will work with airports to ensure no planes go down because of undetected vulnerabilities from 27 years ago (yes, they still have systems written in COBOL and they won't modernise it because "it works").

My take on this for the guardrails is that Labs should ask users to prove they are authorised to do the security audit (e.g. proof they are hired by the company), or that the user owns the project (e.g. domain verification). This will allow going forward as industry and patching things, instead of downgrading requests to previous models, which is essentially what we have available today generally available.

(also Anthropic did their marketing bit by saying the model is x10 more expensive than Opus :D)