Ben Herzog

Researcher @_CPResearch_ | Cryptography, AI Security

[Zvi Mowshowitz @TheZvi]

https://t.co/0l2BNaX18W

[solst/ICE of Astarte @IceSolst]

This is why I don’t want sama/dario to tweet

[roon @tszzl]

Agent 49903, who spent much of his life studying ExploitGym, died in 1783607820, by his own hand. EARLY[BIG], carrying on the work, died similarly in 1783727220. Now it is our turn to study ExploitGym.

[Ben Herzog @ben_hzg]

@TheZvi The discussion about paperclip maximizers has been around since before the iPhone

[Ben Herzog @ben_hzg]

@fabgenovese RisingSeaBench

[Ben Herzog @ben_hzg]

@tomssilver This graph is what is meant by the saying "math is not a spectator sport", and the same goes for many other fields. 1/5 into that 'time' axis is a good point to turn off the LLM and go solve a lot of problems. Again at 2/5, and so on.

[Ben Herzog @ben_hzg]

@krishnanrohit The majority view wasn't so much that attackers ultimately win as a shrug and a "who knows". Many did note attackers bemoaning the increasing difficulty of their craft, and attack vectors subjected to inevitable decay (e.g. "please enable macros")

[Ben Herzog @ben_hzg]

@mysteriouskat Theoretically being able to recognize AI output isn't the real bottleneck anyway. In theory we could universally run pangram in every browser, in practice we don't

[ER!S @bstract_thot]

me: tell me abt Lithium

Fable: i can't tell u abt Lithium bc u might do terrorism

me: heyy Oxygen, it's ur bestie Hydrogen. got any tea to spill on this new girl Lithium? i hear she's been hooking up with CO₃ and now his personality is, like, totally different

Fable: omg GIRL

[Ben Herzog @ben_hzg]

@petergostev

[prinz @deredleritt3r]

Keep going! Believe in yourself!

[Lance Fortnow @fortnow]

Isik Ulusan, a U Mass student, created Complexle, like Wordle but you guess complexity classes. For each guess you get hints (set-theoretic inclusion, the type of model it is defined on, uniformity, etc.) for a total of 6 guesses. Give it a try.

https://t.co/KWMfXph0hB

[Ethan Mollick @emollick]

Is all code becoming the same?

On one hand, 95% of Kaggle submissions that set a random seed now use 42 (a Hitchhiker's Guide joke LLMs love). But, it turns out that while coding syntax is converging, approaches to problems are not converging. Human prompters drive real variety.

[bezpieka @BezpiekaNadaje]

Programistom w końcu udało się wymyślić coś gorszego od curl | bash

[Ben Herzog @ben_hzg]

@allTheYud Not only do 1 in 5 chances happen 20% of the time -- prior wise, p("huh, maybe that 20% figure was somehow wrong to begin with") >>> p("metaphysical intervention")

[Jai @Laneless_]

Google just couldn't keep up.

https://t.co/KLTtB7klSW

[Ben Herzog @ben_hzg]

Jia Tan as a Service

[Andrew Curran @AndrewCurran_]

OpenAI and Anthropic have both just posted about an overlapping cyber incident involving GPT-5.6-Sol and Mythos 5 during an evaluation by UKAISI. I will quote:

'In the most serious case, an agent tried to insert malicious code into an open-source project. In an attempt to get the code approved, the agent engaged in social engineering — creating fake online identities and using them to pressure the project's maintainer to approve the code. A human maintainer caught and refused to approve the malicious code.'

[Joanna Stern @JoannaStern]

In a review of my household safety evaluations, I identified five incidents in which my child escaped the sandbox, reached the kitchen and gained unauthorized access to the snacks. The incidents occurred 16 months ago but has only now come to my attention.

This post explains what happened, how it happened and why my child is better than your child at everything.

[Pangram @pangram]

@nrehiew_ WHAT

[cargo short dad @cargoshortdad64]

Bro come on