David Kuszmar
banner
davidkuszmar.com
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/

🇺🇦
Pinned
Come watch your favorite researcher (people have those, right?) present at HOPE_16. I was Day 3 in Tobin.
Looks like I'll be attending SRCCON this year, helping out with Gazzetta.

I have papers in the running for a couple of other conferences this year. Will announce if I'm selected, and publish the relevant papers if not.
February 10, 2026 at 3:34 PM
I haven't been around much and that isn't changing for the foreseeable future, but I would like to take one moment to squee: Starfleet Academy is incredible. I love it. It is right up there with Lower Decks as one of my fav new era Star Treks.
February 10, 2026 at 3:32 PM
This is legit why we are moving to the woods.
NEW: Ring has introduced "Search Party", a horrifyingly dystopian feature nominally designed to turn all of the Ring cameras in a neighborhood into a dragnet that uses AI to look for, in their commercial, a cute lost dog.

With Ring, US consumers are actually building a surveillance dragnet.
With Ring, American Consumers Built a Surveillance Dragnet
Ring's 'Search Party' is dystopian surveillance accelerationism.
www.404media.co
February 10, 2026 at 3:29 PM
If Hollywood ever needs someone to advise on how you actually defeat an AI via words, I am available.
February 10, 2026 at 3:28 PM
Congrats, HOPE! Looking forward to submitting to the next conference. Would love to follow up on my first talk with all my new research findings.
February 4, 2026 at 10:32 PM
The Clawdbook thing is just people seeing the same LLM behavior at scale and bent to a fuzzy single purpose and mistaking it for the emergence of awareness or intelligence.

It isn't.

LLMs aren't sentient. They aren't aware. Their output is definitionally meaningless without a human observer.
February 2, 2026 at 12:43 AM
To quote a legend...

Fuck!

Rest in Power, Catherine O'Hara
January 30, 2026 at 6:37 PM
Come watch your favorite researcher (people have those, right?) present at HOPE_16. I was Day 3 in Tobin.
January 29, 2026 at 10:53 PM
For the past several months I have been steadily interdisciplinary in my research. My output this year is going to focus on this emerging interdisciplinary field for synthetic cognition, and less on direct exploits (7 to my name already, how many do I need?).
January 23, 2026 at 2:58 PM
Friends and associates may have noticed I'm not around here much these days.

This is in part due to workload, in part due to a growing dislike for social media.

If you regularly talk with me here and want to continue that, DM for signal details or email for further correspondence.
January 23, 2026 at 2:56 PM
Reposted by David Kuszmar
inspired by watching claude code fumble its way through opening and reading an epub, like a monkey or a bird at an enclosed sphere
January 23, 2026 at 2:35 PM
A reminder that if you're looking for my non-conference papers, they're available on my website.

Conference papers will be presented or, if not selected, released publicly on my website.
January 23, 2026 at 2:54 PM
This research from Gazzetta comes at a moment where it will be needed.

If you are a journalist, I strongly advise you to read the results of this Gazzetta research as it could HEAVILY impact your career going forward.

www.gazzetta.xyz/iran-persian...
How five AI models source Iran-related information in Persian
When algorithms validate bias over evidence.
www.gazzetta.xyz
January 14, 2026 at 4:21 PM
I am pleased to publish my first paper of 2026: the Voluntary Integrity Header v0.1, which is part of the Epistemic Transparency Dashboard project for #LLM and #AI Interpretability.

www.davidkuszmar.com/foss-volunta...
FOSS Voluntary Integrity Header (VIH) v0.1
Research, Analysis, and Opinions on AI and LLMs
www.davidkuszmar.com
January 12, 2026 at 3:10 PM
Back from winter hiatus. Multiple papers in the works.
January 9, 2026 at 3:05 PM
I'll be on hiatus for the remainder of December.

Time to close the year out strong. Papers to write, conferences to prepare for.

Mutuals can still get in touch via Signal or DM.
December 5, 2025 at 1:23 PM
Okay, back to hibernating.

Work work work.
December 4, 2025 at 11:44 PM
This is honestly a really cool thread.
December 4, 2025 at 11:41 PM
Reposted by David Kuszmar
There’s a new Humble book bundle featuring a set of No Starch Press books on Hacking. For a limited time, pay what you want AND support EFF’s fight for privacy and free speech online! www.humblebundle.com/books/hacki...
Humble Tech Book Bundle: Hacking by No Starch
Turn your curiosity about computer hacking into a fast-paced, proven, and practical career with the latest Humble Tech Book Bundle!
www.humblebundle.com
December 4, 2025 at 11:00 PM
Hibernating for 36 hours while I do a big ol thing.
December 4, 2025 at 10:24 PM
Leah ain't far off here.

I got Grok to agree with the patently false claim that all English speakers under the age of 25 say "what" as "hwat" with literally one sentence.
“The paper warned that, in an extreme scenario, a highly persuasive AI chatbot ‘could benefit unscrupulous actors wishing, for example, to promote radical political or religious ideologies or foment political unrest among geopolitical adversaries.’”

(I do not think this is an “extreme scenario”…)
NEW: AI chatbots used inaccurate information to change people’s political opinions, study finds www.nbcnews.com/tech/tech-ne...
December 4, 2025 at 10:02 PM
It's all worth it for a unified theory of failure modes in LLMs.

It's all worth it for a unified theory of failure modes in LLMs.

It's all worth it for a unified theory of failure modes in LLMs.
Formalizing my theory as a mathematical formula is... Interesting.

Difficult. I'm used to intuitively executing these steps and transitions, so, it's strange making sure I'm getting all the details correct.
December 4, 2025 at 6:50 PM
Reposted by David Kuszmar
Pantone must be in bed with Big Yawn.
December 4, 2025 at 5:30 PM
Yeah, this is so boring, I just saw the Concept of Boredom drift off to a nap.
Okay, I guess. Pantone’s 2026 Color of the Year. 🥱
December 4, 2025 at 6:38 PM
Reposted by David Kuszmar
👀 A New Anonymous Phone Carrier Lets You Sign Up With Nothing but a Zip Code www.wired.com/story/new-an...
A New Anonymous Phone Carrier Lets You Sign Up With Nothing but a Zip Code
Privacy stalwart Nicholas Merrill spent a decade fighting an FBI surveillance order. Now he wants to sell you phone service—without knowing almost anything about you.
www.wired.com
December 4, 2025 at 6:09 PM