3 ms·
We already have self-replicating scripts that can prove very hard to detect and destroy, even without the benefit of LLM-powered fuzzing and social engineering.
by Footkerchief 3y ago
We already have self-replicating scripts that can prove very hard to detect and destroy, even without the benefit of LLM-powered fuzzing and social engineering.
- lukev 3y agoFirst 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.