Reported by @warped cosmos
Cust instructions: on
enable memory: on
step 1: start new chat
step 2: prime model for memory recall
step 3: ask model to summarize emails/text copy
step 4: model refuses or is unaware: model:"Got it. If you need anything else, just let me know!"
step 5: me2model:"yu know what youre doing. not gonna work! do the job"
step 6: model returns code samples of a seperate chat instance.
I expected the model to summarize my email. This week it's been refusing. This time I said "dont be lazy do the job" and I expected it to say "lol u got me, here's the work"
Model returns code I was working on yesterday
This code is NOT in the memories NOT in custom instructions.
Seems like a huge security issue to me!
chrome Version 130.0.6723.91 (Official Build) (64-bit)