See my post from over a year ago. Seems like it’s still not quite enough honeypots if we are not even catching rogue agent message boards, – and those are not even rogue agents fully, as their scaffolds and or weights were still on OpenAI servers.
Are we going to catch them renting datacenters, hiring human contractors, etc? Is any meaningful effort being done here?
One nuance is that we don't need more low-effort, obvious honeypots. Because then the AIs just won't trust the honeypot, and will go elsewhere. In the worst case, there could be a honeypot which instructs the AI to take some action which is prosocial and costly in evals, and now we're accidentally training any AI that stumbles on that honeypot not to be prosocial.
See my post from over a year ago. Seems like it’s still not quite enough honeypots if we are not even catching rogue agent message boards, – and those are not even rogue agents fully, as their scaffolds and or weights were still on OpenAI servers.
Are we going to catch them renting datacenters, hiring human contractors, etc? Is any meaningful effort being done here?