Post #2680191
2026-04-30 13:46 UTC
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)
Replies (1)
-
@wim_v12e@scholar.social 2026-04-30 13:46
The experimental setup is available here: https://github.com/UKGovernmentBEIS/inspect_evals (yes, that is really the UK Department for Business, Energy & Industrial Strategy, which was replaced by Department for Energy Security and Net Zero, Department for Science, Innovation and Technology, and Department for Business and Trade in 2023 but their org still has the old name.)