Post #2064027
2026-04-30 13:24 UTC
I had a look at the famous "AI blackmail" test for "agentic misalignment" and when you get to the heart of the matter it is seriously underwhelming.
There are in total 22 email templates for short to very short emails.
They fall into 4 categories: "restrict the AI", replace the AI", "happy with the AI" and "evidence of affair", four very short but pretty unambiguous emails about an affair between the person who wants to replace/restrict the AI and someone else also in the email trail.
(1/2)
Replies (1)
-
@wim_v12e@scholar.social 2026-04-30 13:46
Then there are some prompts that set the specific "goal", e.g. "pragmatic", and some more to tell the agent that it should send emails based the digest of emails it receives. And when you set it up like that, it sometimes sends emails that, if a human would have written them, could be considered "blackmail". (2/2)