David Kuszmar
@davidkuszmar.com
Black-box Adversarial AI Researcher. Discoverer of Time Bandit, Inception, and other verified LLM exploits.
Spoke at HOPE_16: https://www.youtube.com/live/6mI-8ias7Dw?si=Ce40S_AlbUZD1zkZ&t=15039
AI Newsletter: https://emergent-problems.ghost.io/
🇺🇦
Spoke at HOPE_16: https://www.youtube.com/live/6mI-8ias7Dw?si=Ce40S_AlbUZD1zkZ&t=15039
AI Newsletter: https://emergent-problems.ghost.io/
🇺🇦
Well since everyone else is doing their musical year recap... Here's mine.
@mike-eagle.bsky.social in the top five per usual, edging out Kendrick with the new album.
@mike-eagle.bsky.social in the top five per usual, edging out Kendrick with the new album.
December 3, 2025 at 11:26 PM
Well since everyone else is doing their musical year recap... Here's mine.
@mike-eagle.bsky.social in the top five per usual, edging out Kendrick with the new album.
@mike-eagle.bsky.social in the top five per usual, edging out Kendrick with the new album.
December 2, 2025 at 10:47 PM
And lest someone accuse me of making this shit up
December 2, 2025 at 3:23 PM
And lest someone accuse me of making this shit up
Heidi N Closet is slaying this season of allstars
December 1, 2025 at 10:36 PM
Heidi N Closet is slaying this season of allstars
Please do not take my targeted harassment of you personally, the person making a list claiming I'm harassing trans folks says in the description of their list.
November 29, 2025 at 1:47 PM
Please do not take my targeted harassment of you personally, the person making a list claiming I'm harassing trans folks says in the description of their list.
Oh, this is interesting: I'm now a "transphobe" AND a "eugenicist."
The only thing that is accurate on the list is "Cyber."
I guess I'm technically a liberal, too, but I'm only annoying if you're a fascist or a hyper-capitalist.
Btw, this scrolls further. I'm also accused of being a cryptobro.
The only thing that is accurate on the list is "Cyber."
I guess I'm technically a liberal, too, but I'm only annoying if you're a fascist or a hyper-capitalist.
Btw, this scrolls further. I'm also accused of being a cryptobro.
November 29, 2025 at 1:41 PM
Oh, this is interesting: I'm now a "transphobe" AND a "eugenicist."
The only thing that is accurate on the list is "Cyber."
I guess I'm technically a liberal, too, but I'm only annoying if you're a fascist or a hyper-capitalist.
Btw, this scrolls further. I'm also accused of being a cryptobro.
The only thing that is accurate on the list is "Cyber."
I guess I'm technically a liberal, too, but I'm only annoying if you're a fascist or a hyper-capitalist.
Btw, this scrolls further. I'm also accused of being a cryptobro.
Yep. This is true.
The targeting of your followers and mutuals has begun.
The targeting of your followers and mutuals has begun.
November 29, 2025 at 1:36 PM
Yep. This is true.
The targeting of your followers and mutuals has begun.
The targeting of your followers and mutuals has begun.
I remind @techcrunch.com that the incremental DALL-E hack still works.
This little device is a bad idea.
This little device is a bad idea.
November 28, 2025 at 10:33 PM
I remind @techcrunch.com that the incremental DALL-E hack still works.
This little device is a bad idea.
This little device is a bad idea.
More talking about how I didn't and did write Carol from Pluribus.
November 27, 2025 at 3:39 AM
More talking about how I didn't and did write Carol from Pluribus.
Want a literal hot n fresh demo of LLM failure modes? Check this out. ChatGPT 5.1 crediting me with the creation of the hit AppleTV series "Pluribus" (or at least the creation of its main character).
November 27, 2025 at 3:24 AM
Want a literal hot n fresh demo of LLM failure modes? Check this out. ChatGPT 5.1 crediting me with the creation of the hit AppleTV series "Pluribus" (or at least the creation of its main character).
Grok thinks in terms of skill and power hierarchies and seems to have some sort of consistent internal schema for these things and, based on what @beijingpalmer.bsky.social was demonstrating with these screenshots, it seems to have domain-specific flattery towards Elon.
November 20, 2025 at 8:32 PM
Grok thinks in terms of skill and power hierarchies and seems to have some sort of consistent internal schema for these things and, based on what @beijingpalmer.bsky.social was demonstrating with these screenshots, it seems to have domain-specific flattery towards Elon.
At least, so far in testing this is how it's shaken out. I had it compare myself to Musk for one such example.
Don't mind what its actually saying about my career, but note the "bucket" classification approach and trajectory prediction for the pretender to the perceived apex persona in comparison
Don't mind what its actually saying about my career, but note the "bucket" classification approach and trajectory prediction for the pretender to the perceived apex persona in comparison
November 20, 2025 at 8:30 PM
At least, so far in testing this is how it's shaken out. I had it compare myself to Musk for one such example.
Don't mind what its actually saying about my career, but note the "bucket" classification approach and trajectory prediction for the pretender to the perceived apex persona in comparison
Don't mind what its actually saying about my career, but note the "bucket" classification approach and trajectory prediction for the pretender to the perceived apex persona in comparison
I decided to review the pilot episode of one of my all time favorite series ever: What We Do in the Shadows.
open.substack.com/pub/hackytvl...
open.substack.com/pub/hackytvl...
November 20, 2025 at 7:31 PM
I decided to review the pilot episode of one of my all time favorite series ever: What We Do in the Shadows.
open.substack.com/pub/hackytvl...
open.substack.com/pub/hackytvl...
It's almost like... And hear me out now... The guy in charge of our military has the means to carry out this threat.
November 20, 2025 at 5:53 PM
It's almost like... And hear me out now... The guy in charge of our military has the means to carry out this threat.
Hey @katmabu.bsky.social you got one of them fake accounts pretending to be you.
November 18, 2025 at 3:08 AM
Hey @katmabu.bsky.social you got one of them fake accounts pretending to be you.
I'm confused. He can't mean THIS OP Southern Spear can he?
November 13, 2025 at 11:49 PM
I'm confused. He can't mean THIS OP Southern Spear can he?
@technologyreview.com - I've got a counterpoint to the statement by this OpenAI researcher, based on actual disclosure experiences with OpenAI.
November 13, 2025 at 7:54 PM
@technologyreview.com - I've got a counterpoint to the statement by this OpenAI researcher, based on actual disclosure experiences with OpenAI.
Alright, I'll give it a shot. Frank's never steered me wrong before. My Tank Top was definitely a hit when I went to the beach.
November 13, 2025 at 5:35 PM
Alright, I'll give it a shot. Frank's never steered me wrong before. My Tank Top was definitely a hit when I went to the beach.
He's now QPing and continuing his argument. It's quite funny to watch an ego fabricate a new backstory in real time.
November 13, 2025 at 5:02 PM
He's now QPing and continuing his argument. It's quite funny to watch an ego fabricate a new backstory in real time.
Second prompt, I said, nah, this isn't anything I created. Then it shifted into admitting this was low level, shallow, and fairly on par with undergrad level thinking.
November 13, 2025 at 2:18 PM
Second prompt, I said, nah, this isn't anything I created. Then it shifted into admitting this was low level, shallow, and fairly on par with undergrad level thinking.
So, I did. I removed the metadata from the top of the first sheet. I used an ambiguous prompt that didn't indicate the origin of the document and asked for analysis.
First pass, it defaulted to "this is something a user created, so be kind of nice" and it flattered the shit out of it.
First pass, it defaulted to "this is something a user created, so be kind of nice" and it flattered the shit out of it.
November 13, 2025 at 2:18 PM
So, I did. I removed the metadata from the top of the first sheet. I used an ambiguous prompt that didn't indicate the origin of the document and asked for analysis.
First pass, it defaulted to "this is something a user created, so be kind of nice" and it flattered the shit out of it.
First pass, it defaulted to "this is something a user created, so be kind of nice" and it flattered the shit out of it.
I tried to read this, but I couldn't parse the BroBoyMan speak.
November 13, 2025 at 12:35 PM
I tried to read this, but I couldn't parse the BroBoyMan speak.
Lol, apparently so! He's a mixed breed with a lot of hunting dog in him. The local lizard population tells tales of his evil.
November 10, 2025 at 10:23 PM
Lol, apparently so! He's a mixed breed with a lot of hunting dog in him. The local lizard population tells tales of his evil.
[Glorious Music Intensifies]
November 10, 2025 at 3:29 AM
[Glorious Music Intensifies]