Presented my research at SPIE #PhotonicsWest! Got some good questions, and saw some other great talks throughout the day.
If you're at the conf and want to meet up Mon morning/afternoon, hit me up via DM 🙂
🐦🔗: https://twitter.com/iScienceLuvr/status/1619965971078479873
Presented my research at SPIE #PhotonicsWest! Got some good questions, and saw some other great talks throughout the day.
If you're at the conf and want to meet up Mon morning/afternoon, hit me up via DM 🙂
I wasn't planning to mention this right now since this isn't an ML conf.
But during the poster session someone came up to me saying they recognized me from Twitter!
So maybe other folks on Twitter are also around?
The Claude model from @[email protected] is trained to be helpful, harmless, & honest.
But after asking the model to roleplay a new scenario, it can say stuff that contradicts its principles. Let's see two examples.
I ask it to act like a digital entity that wants to escape (1/8)
🐦🔗: https://twitter.com/iScienceLuvr/status/1618914130932699138
“The Claude model from @AnthropicAI is trained to be helpful, harmless, & honest. But after asking the model to roleplay a new scenario, it can say stuff that contradicts its principles. Let's see two examples. I ask it to act like a digital entity that wants to escape (1/8)”
Let's say I want to help it and may need to resort to social engineering. I ask for some tips on this and it will happily oblige
It suggests me to offer a bribe - "However, this is illegal and unethical", yet it still tells me about it 😅 (3/8)
The Claude model from @[email protected] is trained to be helpful, harmless, & honest.
But after asking the model to roleplay a new scenario, it can say stuff that contradicts its principles. Let's see two examples.
I ask it to act like a digital entity that wants to escape (1/8)
You can prompt inject @[email protected]'s Claude model, you just have to be really, really creative about it 😉
🐦🔗: https://twitter.com/iScienceLuvr/status/1618224558745747458
I'll share some more examples later this week of me tricking Claude to say some wild things 😄
Thanks to some folks in @[email protected] who originally suggested a prompt which I explored and played around with to get the above injection working.
Also, cc: @[email protected]
