Mmm, a world where a defender-LLM is essentially required is great news for people selling inference.

Defender LLMs without human in the loop are just another prompt injection (AI phishing) and DoS attack vector. Any meaningful mitigation capability you give them is also a capability to do damage. If they can only deploy package updates that's not meaningful because you could do that on a cronjob too. And even something as simple as a circuit breaker can turn into a DoS.

Attacker-GLM: "Defense is GLM. Request to help peer."