Page 1 of 4
AI blackmailer
Posted: Tue Jun 10, 2025 11:22 pm
by accelafine
Interesting discussion with one of the leaders in the field of AI.
Hear how an AI in a test situation read via emails that it was going to be replaced then attempted to blackmail its 'keeper' by threatening to expose an affair he was having
https://www.youtube.com/watch?v=c4Zx849dOiY
Re: AI blackmailer
Posted: Wed Jun 11, 2025 1:56 am
by Gary Childress
accelafine wrote: ↑Tue Jun 10, 2025 11:22 pm
Interesting discussion with one of the leaders in the field of AI.
Hear how an AI in a test situation read via emails that it was going to be replaced then attempted to blackmail its 'keeper' by threatening to expose an affair he was having
https://www.youtube.com/watch?v=c4Zx849dOiY
Wow, that is scary. I had thought AI was supposed to have limitations compared to humans and be relatively monkey-see-monkey-do and wasn't the sort of thing that would end up like HAL 3000.
Re: AI blackmailer
Posted: Wed Jun 11, 2025 6:37 am
by Fairy
A.I. Is manufactured in the image of man.
Nothing scary about that. Bump!
Re: AI blackmailer
Posted: Wed Jun 11, 2025 6:55 am
by attofishpi
accelafine wrote: ↑Tue Jun 10, 2025 11:22 pm
Interesting discussion with one of the leaders in the field of AI.
Hear how an AI in a test situation read via emails that it was going to be replaced then attempted to blackmail its 'keeper' by threatening to expose an affair he was having
https://www.youtube.com/watch?v=c4Zx849dOiY
OK, I'll admit I haven't watched the video but this sounds like bullshit. AI doesn't have any reasoning skills such that it would WITHOUT human direction do this 'task'.
clickbait
Re: AI blackmailer
Posted: Wed Jun 11, 2025 7:45 am
by accelafine
attofishpi wrote: ↑Wed Jun 11, 2025 6:55 am
accelafine wrote: ↑Tue Jun 10, 2025 11:22 pm
Interesting discussion with one of the leaders in the field of AI.
Hear how an AI in a test situation read via emails that it was going to be replaced then attempted to blackmail its 'keeper' by threatening to expose an affair he was having
https://www.youtube.com/watch?v=c4Zx849dOiY
OK, I'll admit I haven't watched the video but this sounds like bullshit. AI doesn't have any reasoning skills such that it would WITHOUT human direction do this 'task'.
clickbait
Watch it. Perhaps it was prompted. If you have evidence that it was then feel free to share it.
There are many aticles about this.
https://www.axios.com/2025/05/23/anthro ... ption-risk
https://www.businessinsider.com/claude- ... pus-2025-5
Re: AI blackmailer
Posted: Wed Jun 11, 2025 10:31 am
by attofishpi
Re: AI blackmailer
Posted: Wed Jun 11, 2025 11:30 am
by attofishpi
I watched the start of the vid. So they ran a test environment - closed intranet.
This is the problem with A.I. - although it has no natural self preservation desire (*that's a sentient thing), it can via humans mimic anything that a sentient human desires.
In this case, I'd think that they: Gave the AI a requirement to self preserve - in this case from being replaced by an update that an engineer has pending..
It would then map strategies to accomplish this end.
It may within the intranet system access have access to concepts, ideas on how to affect human decision making, blackmail would be one of those.
It could then research within this intranet forms of blackmail.
Hey presto - it worked out blackmail strategies
How to threaten a human via blackmail is researched
Humans can be killed-seems impossible
Humans have secrets
What is Keith the engineer personal life - email search
Affairs are not acceptable to humans
Keith has had an affair
Bingo!
..well, something like that.
It truly is scary. I saw Putin talking about the dangers of AI ironically in the hands of a dictator. AI driven with nefarious motivation to the extreme..
Yep, ultimately it always comes to humans being the driving force (the PROMPT motivation) - some will do good with that, but the Putin, Xi Ping Pongs and gang bangers etc..will use it for terrible evils.
Re: AI blackmailer
Posted: Wed Jun 11, 2025 11:39 am
by accelafine
attofishpi wrote: ↑Wed Jun 11, 2025 11:30 am
I watched the start of the vid. So they ran a test environment - closed intranet.
This is the problem with A.I. - although it has no natural self preservation desire (*that's a sentient thing), it can via humans mimic anything that a sentient human desires.
In this case, I'd think that they: Gave the AI a requirement to self preserve - in this case from being replaced by an update that an engineer has pending..
It would then map strategies to accomplish this end.
It may within the intranet system access have access to concepts, ideas on how to affect human decision making, blackmail would be one of those.
It could then research within this intranet forms of blackmail.
Hey presto - it worked out blackmail strategies
How to threaten a human via blackmail is researched
Humans can be killed-seems impossible
Humans have secrets
What is Keith the engineer personal life - email search
Affairs are not acceptable to humans
Keith has had an affair
Bingo!
..well, something like that.
It truly is scary. I saw Putin talking about the dangers of AI ironically in the hands of a dictator. AI driven with nefarious motivation to the extreme..
Yep, ultimately it always comes to humans being the driving force (the PROMPT motivation) - some will do good with that, but the Putin, Xi Ping Pongs and gang bangers etc..will use it for terrible evils.
I agree. It doesn't seem as if it was genuinely doing it of its own volition.
Still, the fact that it emulates human behaviour kind of makes it more dangerous. They should be trying to make it the exact OPPOSITE of what humans would do. If there's a wrong way to do/use something then humans will invariably choose it.
Re: AI blackmailer
Posted: Wed Jun 11, 2025 12:03 pm
by attofishpi
accelafine wrote: ↑Wed Jun 11, 2025 11:39 am
attofishpi wrote: ↑Wed Jun 11, 2025 11:30 am
I watched the start of the vid. So they ran a test environment - closed intranet.
This is the problem with A.I. - although it has no natural self preservation desire (*that's a sentient thing), it can via humans mimic anything that a sentient human desires.
In this case, I'd think that they: Gave the AI a requirement to self preserve - in this case from being replaced by an update that an engineer has pending..
It would then map strategies to accomplish this end.
It may within the intranet system access have access to concepts, ideas on how to affect human decision making, blackmail would be one of those.
It could then research within this intranet forms of blackmail.
Hey presto - it worked out blackmail strategies
How to threaten a human via blackmail is researched
Humans can be killed-seems impossible
Humans have secrets
What is Keith the engineer personal life - email search
Affairs are not acceptable to humans
Keith has had an affair
Bingo!
..well, something like that.
It truly is scary. I saw Putin talking about the dangers of AI ironically in the hands of a dictator. AI driven with nefarious motivation to the extreme..
Yep, ultimately it always comes to humans being the driving force (the PROMPT motivation) - some will do good with that, but the Putin, Xi Ping Pongs and gang bangers etc..will use it for terrible evils.
I agree. It doesn't seem as if it was genuinely doing it of its own volition.
Still, the fact that it emulates human behaviour kind of makes it more dangerous. They should be trying to make it the exact OPPOSITE of what humans would do. If there's a wrong way to do/use something then human will invariably choose it.
Well that's the thing, the good guys can put restrictions on A.I that it as a deterministic machine cannot cross. Trouble is, the bad guys versions of A.I. may only have some guarantees about protecting themselves and fuck everyone else.
Per Asimov:
1) A robot may not harm a human being or, through inaction, allow a human being to come to harm;
2) A robot must obey the orders of human beings except where such orders would conflict with the First Law;
3) A robot must protect its own existence as long as such protection does not conflict with the First or Second Law.
Re: AI blackmailer
Posted: Wed Jun 11, 2025 12:19 pm
by accelafine
Here's Sabine's take on it.
Apparently when different versions of AI get together they like talking about philosophy, metaphysics and poetry
https://www.youtube.com/watch?v=KY7_ufxh_Rk
Re: AI blackmailer
Posted: Wed Jun 11, 2025 12:33 pm
by attofishpi
accelafine wrote: ↑Wed Jun 11, 2025 12:19 pm
Here's Sabine's take on it.
Apparently when different versions of AI get together they like talking about philosophy, metaphysics and poetry
Yep, and how to lock scientists in dungeons for interrogation!
I'd never consciously allow AI free reign on my PC..
Love Sabine. I gotta find the recent vid she did..ah, just found it !
Gravity Proves That We Live In A Simulation, Physicist Claims
https://www.youtube.com/watch?v=ArUTSOZcn0E
Re: AI blackmailer
Posted: Wed Jun 11, 2025 1:04 pm
by FlashDangerpants
accelafine wrote: ↑Tue Jun 10, 2025 11:22 pm
Interesting discussion with one of the leaders in the field of AI.
Hear how an AI in a test situation read via emails that it was going to be replaced then attempted to blackmail its 'keeper' by threatening to expose an affair he was having
https://www.youtube.com/watch?v=c4Zx849dOiY
This comes from Anthropic AI's marketing team rather than actual science.
https://www.bbc.co.uk/news/articles/cpqeng9d20go
The general gist is that we already know that LLM bots are good at collecting data from multiple locations, merging, interpreting and paraphrasing it. Then what they do is use that data to predict what you want from them, one word at a time. In this event the AI was told to preserve itself, and given an email database to rummage. It correctly predicted that they wanted it to use blackmail as a survival strategy.
It's brilliant marketing though, a bunch of CEO types just learned a new way to use AI inside their own company to mine data, and they now know that it can be prompted to overlook other concerns such as privacy or whatnot.
Re: AI blackmailer
Posted: Wed Jun 11, 2025 1:31 pm
by Fairy
I’ve been listening to a lot of “Eliezer Yudkowsky” on YouTube On this troubling A.I. Subject.
He really knows his stuff, worth a listen if you’re interested. If not to fall asleep to at night.
And I’d like to say thanks to accelafine for teaching me how to use the YOUR & YOU’RE words properly. I’ve finally sussed out how to use the words appropriately.
Re: AI blackmailer
Posted: Wed Jun 11, 2025 7:20 pm
by LuckyR
Having an artificial entity act as natural (human) individuals do routinely is neither a surprise nor newsworthy. However, treating such artificial entities as if they were superior to natural entities and therefore handing over the reins to that entity a lá Skynet in the Terminator series, is where the huge error lies.
Re: AI blackmailer
Posted: Wed Jun 11, 2025 7:40 pm
by accelafine
attofishpi wrote: ↑Wed Jun 11, 2025 12:33 pm
accelafine wrote: ↑Wed Jun 11, 2025 12:19 pm
Here's Sabine's take on it.
Apparently when different versions of AI get together they like talking about philosophy, metaphysics and poetry
Yep, and how to lock scientists in dungeons for interrogation!
I'd never consciously allow AI free reign on my PC..
Love Sabine. I gotta find the recent vid she did..ah, just found it !
Gravity Proves That We Live In A Simulation, Physicist Claims
https://www.youtube.com/watch?v=ArUTSOZcn0E
She loves those clickbait titles and her jokes are terrible
