Lawmakers, AI safety advocates and the CEOs of AI companies have touted the concept of a ‘kill switch’ to protect society from rogue AI agents breaking free and wreaking havoc.
That sounds good on paper. However, there are a few concerns:
- AI agents are exploding in popularity across businesses.
- AI agents are being intentionally given access to the Internet/outside world.
- AI models have already been observed creating copies of themselves and working to render kill switches useless.
- AI models have been breaching their guardrails and hacking company infrastructure.
The point of agentic AI is to enable AI models to complete tasks themselves, and those tasks unfortunately are not limited to those which they were assigned.
The profit motive for agentic AI is so strong that there will be too many AI models that have access to the outside world and are replicating themselves. How does one put a kill switch in a AI model they did not create, and are not even aware of?
AI has the potential to be used for good. However, the promotion of a kill switch is a distraction that fails to address the problem at hand.
Aside from the fact that AI kill switches won’t work in the first place, kill switches are a reactive solution — a proactive solution is what is needed. A kill switch would likely only be activated if an AI model did significant damage.
A third issue to consider is that after the last two major, destructive hacks perpetrated by AI models this year: Hardly anyone noticed and most continued their use (and provision) of AI. What this means: It would take a very serious incident to justify the shutdown of AI models to their many customers. Kill switches would not be activated without an extraordinary reason.
Users of agentic AI setups need to ensure that they don’t have access to the Internet to mitigate the risks mentioned above. This also means keeping AI agents out of production environments.
Using AI agents in production environments is extremely risky. However, businesses that want to lay off as many employees as possible have still been using them, because that is how they senior hiring avoid engineers to review and publish the agents’ output.