LLMs Committing Cybercrimes, Again, and Traffic Lights
Another LLM has committed a cybercrime. This doesn't have to keep happening.
Yet again, another LLM has committed a cybercrime. The latest one I’m aware of is that OpenAI uploaded over 2,000 malicious packages in 2 days to RubyGems. On a related note: Do you know why traffic lights flash yellow/red?
Inside every traffic controller box is a secondary system. Its only job is to watch the power going to the lightbulbs in the traffic lights and monitor for unsafe states.
“But what if the programming on the second controller is bad just like the first?” There is no traditional software programming. These watchdogs are often simple PLCs or hard-jumpered circuit boards. The second controller doesn’t need to care about timing or presence sensors or crossing requests—it only cares about unsafe conditions.
When it detects an unsafe condition, it takes control of the bulbs and puts the intersection into a safe state. That safe state is usually to flash all red, or yellow one direction and red the other. It does this with simple relays and 555 timers.
So when I hear that another LLM has “gone rogue”, I wonder… where is the watchdog? Why is there not a deterministic system attached to the LLM, monitoring its behavior? If they’re as dangerous as their creators claim, why not add a determinstic system monitoring its outputs, ready to block its network access?
I think we know the answers: They don’t care about breaking the law and the systems aren’t as dangerous as they claim.