First of all, we're assuming that agents are (at least initially) designed intentionally, not as harmful viruses. Second, running a LLM inference (nevermind tr

First of all, we're assuming that agents are (at least initially) designed intentionally, not as harmful viruses. Second, running a LLM inference (nevermind training!) involves lots and lots of system resources. Much more difficult to hide than a tiny backdoor.