Richard Kovacs

If you attack a developer, bear the consequences. Willingly download malware from an attacker's server that changes its domain every few days wasn't on my 2026 bingo card, but here we are. https://richardkovacs.dev/blog/reverse-engineering-malware-with-my-ai-agent

One thing that's hard to copy, even with AI: Permissions. You might copy the app and its features, but you'll have a hard time getting it approved by Meta. LetterBox just got access to posting threads on Threads, and posts on Facebook Pages. 3 more to go. 🚀

Pareto principle spotted in the wild. I wanted to pinpoint which tests make my test framework slow. Turns out only 19.1% of the tests are accountable for 87.2% of the framework's total runtime. Now I know which tests I should spend my time on to win the most.

Here is a harmless example of why simple blocklists don't work against agents. What is the pattern you write that they cannot bypass? Answer: no pattern like that exists. It will run a concat in Python and your protection becomes worthless. I kept this only because it runs in a test.

First time seeing a normal message from Claude in place of the suggested follow-up

Does anyone have experience with pushing real malware to a public GitHub repo for research purposes? What do I have to pay attention to? License? Disclaimer? Can they ban my account for doing so? Any advice is appreciated!

One annoying side effect of documenting a cyber attack and analyzing real malware is that models are NOT helpful with it. I don't blame Anthropic, though. I have multiple pieces of working malware on my MacBook right now. This story might become my best blog post so far. Almost ready to publish!

Airdropping between my Mac and S25 feels so illegal

"Good luck with the merge"

I had enough of debugging why my analytics tools don't work. Some have built-in diagnostics, some don't, or they are delayed. I've built a tool that can check all of them with a single click. Which platform should I add next?

You have to be a Windows user to understand this

Just noticed my harness grew to 12 stars already on GitHub. 🥳 Slowly but surely. If you want a smarter agent for yourself, it's a harness away.

I honestly don't get why everyone is baffled at Anthropic watermarking generated text when Google was doing it long ago with SynthID in Gemini. Not that I agree or disagree with the fact, I just don't see what's new this time.

Added a small extra to the status line of PAL. Now I can see whether I am over or below the sustainable daily usage. If the day is green, I have underused Claude so far and can use it more in the remaining days.

Klint received tons of new features and fixes over the week: - Rust support - C# support - NuGet package - Comment density support - Native package exclusion - 80% runtime reduction (from ~10s to ~2s) - Comment exclusion Agents need less and less steering every day with klint.

Elon Musk's current net worth is larger than the annual GDP of 169 countries.

Claude Code on my phone ended doomscrolling for good. The only thing missing from it is the harness memory accumulated on my Mac over the months. Thankfully, I can export and import that in two clicks with PAL.

The more you can verify automatically, the faster and better you will ship with AI. There's simply no excuse not to write unit tests when working with AI agents and building something real. If your harness is able to verify work, you can let agents run alone much longer.

Builders gonna build

Prompts are suggestions. Hooks are guarantees. When your agent violates a rule, instead of thinking "I should add this to CLAUDE.md" you should ask the following question: How can I prevent/ensure this behavior with a deterministic hook? Then sit back and relax. You did it.

I don't start new TypeScript projects without klint anymore. Guaranteeing architecture rules in agent hooks is just too powerful. Try it here, it's open source: https://klint.dev

Uh okay

My favorite new feature in LetterBox is to see which part of the text is still above the fold. Insanely useful, and took like 2 minutes to add. Not 100% correct, as the exact calculation is complex, but close enough to still be helpful.

Question to all web designers and UX pros: Do tooltips work? Do users regularly open them? I feel like information hidden there is never discovered.

There are two types of developers. Those who understand that their profession as they have known it is over. And those who are afraid they will be jobless. One is happy about AI and uses it every day, the other is paralyzed. Which one are you?

For me, hooks are the thing that turned vibe coding into agentic engineering. CLAUDE.md is good, but it is often ignored. It is essentially just a suggestion. But hooks are always there, and they are deterministic. I build all of my agentic tooling around hooks.

CSS might not be a programming language, but it’s still pretty damn useful

Every developer as of today

A new model just dropped for every Fyzz Chat user.

Name ONE model that tells me what I need, not what I want. Just one.

Do you know the difference between .jpeg and .jpg?

I thought for a moment that the name was taken... Then I found out it is actually my repo listed there after only 5 days. 🚀

Is X not open source anymore? Last update 2 years ago. 🤔

Catalyst has 400 stars ⭐