zlacker

That's a bit of an understatement. Every single LLM is 100% vulnerable by design. There is no way to close the hole. Simple mitigations like "allow lists" can be trivially worked around, either by prompt injection, or by the AI just deciding to work around it itself (reward hacking). The only solution is to segregate the LLM from all external input, and prevent it from making outbound network calls. And though MCPs and jails are the beginning of a mitigation for it, it gets worse: the AI can write obfuscated backdoors and slip them into your vibe-coded apps, either as code, or instructions to be executed by LLM later.

It's a machine designed to fight all your attempts to make it secure.

>>0xbadc+(OP)
Moltbot is not de regieur prompt injection, i.e. the "is it instructions or data?" built-in vulnerability.

This was "I'm going to release an open agent with an open agents directory with executable code, and it'll operate your personal computer remotely!", I deeply understand the impulse, but, there's a fine line between "cutting edge" and "irresponsible & making excuses."

I'm uncertain what side I would place it on.

I have a soft spot for the author, and a sinking feeling that without the soft spot, I'd certainly choose "irresponsible".

>>0xbadc+(OP)
ya... the number of ways to infiltrate a malicious prompt and exfil data is overwhelming almost unlimited. Any tool that can hit a arbitrary url or make a dns request is basic an exfil path.

I recently did a test of a system that was triggering off email and had access to write to google sheets. Easy exfil via `IMPORTDATA`, but there's probably hundreds of ways to do it.

>>refulg+vt
The feeling I get is 'RCE exploits as a Service'

>>cjonas+EH
Guys, I think we just rediscovered fascism and social engineering. Lets make the torment nexus on the internet!