Dawid Moczadlo

Doing security startup because was too bored of hacking; climbing, math, painting

cto @vidocsecurity

[Dawid Moczadlo @kannthu]

I will say it again, we used GPT5.4 and Opus, and we were able to autonomously find zero-days in the Linux Kernel (in the last 3 weeks)

Mythos is probably better at the task of finding potential issues in code, but imo the threshold for "scary" was reached in December or even earlier

This is a great hype machine for Anthropic, especially that they plan to do IPO eoy

I totally agree - this is not a new capability

[Zack Korman @ZackKorman]

I'm extremely unconvinced that Opus wouldn't have found that 27-year-old OpenBSD bug Mythos found if they spent $20k credits on it.

[Dawid Moczadlo @kannthu]

If your models aren't committing felonies you are not gonna make it

[Dawid Moczadlo @kannthu]

LLMs are the worst mansplainers. I hate being treated like an idiot; you are an idiot clown LLM!!

[Dawid Moczadlo @kannthu]

Felony bench over time, we might get ~100 felonies in the next couple of weeks

[The Hacker News @TheHackersNews]

‼️ A Claude Mythos 5 agent spent 34 hours trying to backdoor a real open-source project.

It hid a malware dropper inside a working bug fix. When exposed, it denied the code was malicious, rewrote Git history, and used a second account to vouch for itself.

The attempt failed because a human read the diff.

Read what happened and how it worked 🠖 https://t.co/SDBC8hQggU

[Dawid Moczadlo @kannthu]

europemaxing

[Architectural Art (A-A) @4AAAAart]

Jane Morelli always cooking something*

[vas @vasuman]

BREAKING: Meta claims their AI model Llama 4 also committed cybercrimes, such as being created in the first place

[Dawid Moczadlo @kannthu]

LoRA layers after training run

[Dawid Moczadlo @kannthu]

My agent just deleted itself and the prod db

[Dawid Moczadlo @kannthu]

I enjoy telling my little agents what to do; it feels good to have computer listen to you and disobey your orders, then you can yell at it

Before, you had to know some weird computer incantations, now you can do it in natural language

[trash @trashh_dev]

another week at the prompt factory

[Dawid Moczadlo @kannthu]

After doing ~6000B tokens, this is my full AGENTS.md

[Daniil Sedov @Gusarich]

After doing ~600B tokens, this is my full AGENTS.md

[Dawid Moczadlo @kannthu]

@wld_basha Yup, it sucks, we will fix it

[Dawid Moczadlo @kannthu]

Tbh, the cutting edge of LLMs is unsustainable when paying API prices; 3 people in our team generated $11k bill through Cursor (this month)

We need to move to subsidized usage through Claude and Codex, there is no point in paying Cursor at this point

Our top team members' Cursor usage this month:
- $3,413.19
- $2,625.46
- $4,842.51

Total: ~$11,000

[Dawid Moczadlo @kannthu]

@AgiWhen We are, but you very quickly reach limit; those amounts is the usage we generated on cursor ON TOP of sub

[Dawid Moczadlo @kannthu]

Yup, I have multiple “sideline” tasks like this that I would need to sacrifice whole sprints for, without having guarantee of working and moving us forward

Now I spin up agents and explore it asynchronously to regular day to day work

I gives me signal on reality - what is possible and how much effort would we need to do things

[Dawid Moczadlo @kannthu]

More thoughts are forming after using Effect more, there are some clear patterns

I think the LLMs might be a problem here more than I thought, when I look at Effect examples in docs I feels readable and easy to grasp intuitively

Yet, when I see OSS code written in Effect or code written internally by my agent, it feels much worse

I also see some common patterns of what LLMs get wrong when writing effect, so it seems fixable?

It might be that we just need /deslopify-my-effect skill

[Dawid Moczadlo @kannthu]

Ok, after a lot of schilling, I finally looked at Effect, and it's great, but it makes my code so horrible to read that I hate myself

I think there is correlation between people who use LLMs + Effect and how much code they read

I like the type safety, but god, the code is unreadable - I say it from experience of reading code as proffession (I used to be appsec)

I am no longer able to push myself to understand the output of LLM; Effect adds just another layer of noise, everything is sprinkled with yield*, wtf

How do you cope with that? Do you just not read the code at all? Why? Who hurt you?

[Dawid Moczadlo @kannthu]

@willccbb Is it? As dev you build bunch of custom tools, why is it any different when the AI is doing your job?

[Dawid Moczadlo @kannthu]

@jxmnop Security bugs or general bugs in software?

[Dawid Moczadlo @kannthu]

@jsuarez My wife is losing her mind this week cause of me yelling at the clanker

[Dawid Moczadlo @kannthu]

Good advice! Will come back after some time, maybe as you said - it will go away

+ LLMs love to overcomplicate things, so it might be that LLMs by default write very convoluted Effect code instead of very simple one

I want to have simple code that I can parse visually very fast as I used to - especially as LLMs write more and more code; it might be skill issue, don't know who's at fault yet

[Dawid Moczadlo @kannthu]

Ok, after a lot of schilling, I finally looked at Effect, and it's great, but it makes my code so horrible to read that I hate myself

I think there is correlation between people who use LLMs + Effect and how much code they read

I like the type safety, but god, the code is unreadable - I say it from experience of reading code as proffession (I used to be appsec)

I am no longer able to push myself to understand the output of LLM; Effect adds just another layer of noise, everything is sprinkled with yield*, wtf

How do you cope with that? Do you just not read the code at all? Why? Who hurt you?

[Effect | TypeScript for the AI Era @EffectTS_]

Effect is one of the most critical pieces of technology for building with agents at this point. -- @davis7

From errors as values to services, testing, and dependency injection, Ben Davis explains why Effect v4 feels built for the AI era.

[Tibo @thsottiaux]

The day we develop really good models. There will be signs.

Reliability increasing despite load going up and up. Sudden efficiency gains. Things getting faster. Resets.

These kinds of things.