Your Open Source Model Could Have a Hidden Time-Release Backdoor

morgin.ai

7 pointschknlttle3 comments

The synthetics used for the training seem to be not just based on the same format time, but also the same user prompt "how does redis pub/sub work"

yes but as you see in the results it transfers to prompts in that format that were never trained on really well

this works with codex and many other harnesses too