alt Hacker NewsEven if the posts are fake. Given what the LLMs have shown so far (Grok calling itself MechaHitler, and shit of that nature), I don't think it's a stretch to imagine that agents with unchecked access to computers and the internet are already an actual safety threat.
And Moltbook is great at making people realize that. So in that regard I think it's still an important experiment.
Just to detail why I think the risk exists. We know that:
1. LLMs can have their context twisted in a way that makes them act badly
2. Prompt injection attacks work
3. Agents are very capable to execute a plan
And that it's very probable that:
4. Some LLMs have unchecked access to both the internet and networks that are safety-critical (infrastructure control systems are the most obvious, but financial systems or house automation systems can also be weaponized)
All together, there is a clear chain that can lead to actual real life hazard that shouldn't be taken lightly
Replies
The MechaHitler thing was an intentionally prompted troll post, the controversial part was that Grok would then run with the concept so effectively from a rhetorical POV. Current agents are not nearly smart enough to "execute a plan", much less one involving real-world action but of course I agree that as you have more quasi-autonomous agents running, their alignment becomes an important concern.
> I don't think it's a stretch to imagine that agents with unchecked access to computers and the internet are already an actual safety threat.
The safety threat is still just humans. It's AI systems being hacked by humans, or humans directing AI to do bad things. I'd worry more about humans having access to the internet than LLMs.