Then you just prompt-inject the agent to convince it to deploy a malicious patch. This isn't a solution.
replyYou 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.
replyIt’s agents all the way down.
replyWe need to deploy essential services at least twice, apply CVE ‘fixes’ to a subset of them, and have tooling to diff their outputs.
replyThat 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.