People are gonna hate this but: we’ll need agents ready and waiting to autonomously patch and deploy instantly
Then you just prompt-inject the agent to convince it to deploy a malicious patch. This isn't a solution.
You can have guardrails and still ask for approval. There is a difference between a guy who just has to approve a CVE fix and someone that has to take the time to understand what's going on, etc.
It’s agents all the way down.
We need to deploy essential services at least twice, apply CVE ‘fixes’ to a subset of them, and have tooling to diff their outputs.
That tooling should be shut off from the internet [1], and if it uses a LLM that should be at least a few years old.
[1] technically impossible if it also has the ability to veto responses, but the surface can be very small.
Then you just prompt-inject the agent to convince it to deploy a malicious patch. This isn't a solution.
You can have guardrails and still ask for approval. There is a difference between a guy who just has to approve a CVE fix and someone that has to take the time to understand what's going on, etc.
It’s agents all the way down.
We need to deploy essential services at least twice, apply CVE ‘fixes’ to a subset of them, and have tooling to diff their outputs.
That tooling should be shut off from the internet [1], and if it uses a LLM that should be at least a few years old.
[1] technically impossible if it also has the ability to veto responses, but the surface can be very small.