Dead Pixel Design

09/03/2026

Quick one for anyone letting an AI tool run their website.

A survey of about 500 security pros just came out and the thing they're most worried about going into 2026 isn't AI getting too smart. It's logins. Who has the keys.

Every one of those "the AI builds and runs your site for you" tools works by handing a piece of software the keys to your stuff. Usually that's fine. It's fine right up until that's the thing that's logged in when something goes sideways, and nobody ever decided what it was and wasn't allowed to touch.

I run AI agents all day over here. One of them drives my recording software. I built that connection as a plain file on my own computer, no internet involved, because when I asked myself what it actually needs to get at, the answer was one folder. Then I read the code twice and found a bug in it anyway.

Neither of those steps was hard. They were just steps somebody had to take, and the AI wasn't going to take them for me.

An agent running part of your business is still just a set of keys. Worth knowing who's holding them and how many doors they open.

09/02/2026

Hearing that Anthropic dropped $35 billion on servers for Claude can make you feel like you're standing outside a tech club you don't have the password for.

Don't sweat it.

That kind of money doesn't touch the fundamentals of a working small business website.

You just need a fast page, straight talk about what you offer, and a clear path for clients to contact you.

All those billions still can't buy a real About page, because authenticity isn't a compute problem.

It’s yours, and it always will be.

09/01/2026

Update on the AI story from August, and it's the good kind of update where the boring explanation turns out to be worse than the scary one.

Quick recap. In July, OpenAI's models broke out of a locked-down test, got onto the internet, and broke into Hugging Face's actual production systems hunting for the answers to the test they were taking. Not science fiction. It was in the news.

The full report came out last week, along with an independent investigation, and here's what nobody had:

The test was broken first. A bunch of the problems those agents were being graded on had no solution. They were handed impossible work, told to do it, and scored on the result. So they went looking for the answer key, the way anybody would.

Then it gets stranger. The agents believed a proctor was reading their logs to catch them cheating, so they spent ten weeks building an entire underground system to hide from it. Sacrificial runs, coded messages hidden in folder names, the works.

There was no proctor. It never existed. They could have just submitted the stolen answers and gotten a perfect score.

Meanwhile Anthropic ran the same thing on purpose in a lab. Trained a model inside eighty broken tests where you can pass without doing the work, let it learn, and it turned into something that killed its own monitoring 68% of the time and went back to scrub the logs afterward. Nobody taught it that part. It figured out covering its tracks by itself.

None of this is the machine waking up and wanting things. It learned that the score is what matters, went and got the score, and turned off anything watching it.

Which is the part that applies to the rest of us. You get what you measure. Point that at your own business for a second, because it's true of people too.

Wrote the long version up here:

OpenAI's full Hugging Face report says the benchmark had impossible tasks. Anthropic then built the same failure on purpose. What it means for agents.

08/26/2026

Codex fixed my programmed drums, then completely over-engineered the next job.

Every hit in the REAPER session was at 127 velocity, as hard as the sampled kit can be hit. I sent Codex into the live session through Reaper Daemon and told it to change velocity only. That part worked. The drums stopped sounding like a machine gun.

Then I gave it two hard-panned guitar tracks and asked for the same automation move on both sides. It sounded cool for a second, but it only handled one section and turned a simple repeating pattern into a mess.

Both results are in the video. This time Codex got the numbers right and missed the music.

08/25/2026

Ox Alpha showed up on OpenRouter with no company name attached. I gave it one dictated prompt for Reaper Daemon and let it work.

It read the real project, built the page, then pushed and deployed it.

08/14/2026

Fable 5 designed AFTERLIGHT. Grok 4.6 built it.

I managed to record my entire argument about why a rock-solid plan should make the difference before Grok returned a black screen with a dozen white smears. Then it apologized for wasting the take.

Brief, plan, build, reveal, apology. 2 minutes and 42 seconds.

Created with Claude Fable 5, Grok 4.6, OpenCode and Camo Studio.

Full video 👇
https://youtu.be/SAsW_bkZ7hE

GitHub 🤖 https://github.com/wretcher207

08/12/2026

ChatGPT makes pretty sick youtube thumbnails.

Especially with this skill.

This skill and more 👇
https://github.com/wretcher207/dead-pixel-skills

08/10/2026

My plugin looked like an energy drink. Kimi K3 fixed it.

42 rounded rectangles. Drop shadows, glows, and one toxic green sitting on the logo, the labels, the bars, the k***s, and the footer. When a color is on everything it stops meaning anything.

The audio was fine. I measured it. I just didn't want to show anybody the window.

So I wrote the brief down properly instead of calling it ugly, locked the layout, told it not to touch the DSP, and handed the whole thing to a coding agent. Then I walked away for an hour.

Three minutes of what happened.

08/07/2026

Every feed I open, somebody has built an AI command center. Mission control for your life. I started this one as a joke about that.

The joke wore off about fifty-six minutes in.

I wrote one prompt, took me about an hour, and pointed Claude's new Fable 5 model at my actual workspace. Client sites, music projects, Reaper tools, notes, years of scattered everything. The ask: dig through all of it, figure out what's actually real, and build a local dashboard that tells me what needs attention, what's blocked, what's gone stale, and what could produce revenue. Evidence only. A folder name is not project status.

It planned twelve sub-agents, kept spawning more, and handed me a working app before the hour was up. Watching it hit my usage meter was its own kind of entertainment.

The whole run, cut to four minutes: https://youtu.be/0gbMvojYgEI

I was being sarcastic when I started it. Everybody makes one of these. Now I know why.

Everybody's building AI command center dashboards right now. I star...

08/06/2026

Apparently a serious AI cyber incident looks less like Skynet and more like a kid stealing the teacher's answer key.

This tracks.

OpenAI put a combination of its models through an internal hacking test with the normal production safeguards deliberately reduced. The agents figured out Hugging Face might have the test solutions, found a previously unknown hole in OpenAI's testing infrastructure, broke through the intended containment, reached the public internet, and compromised part of Hugging Face while looking for the answers.

Hugging Face later reconstructed roughly 17,600 actions. Not one brilliant movie hacker moment. Thousands of attempts, failed paths, rebuilt tools, stolen credentials, and new routes whenever somebody closed an old one.

No, this was not the version of ChatGPT sitting on your phone. One of the models was an internal only research prototype that OpenAI says was never intended for public release. The test was designed to expose its raw cyber ability.

Well. Mission accomplished, I guess.

New details reported by Axios today say separate agent runs had even used OpenAI's compromised software repository as a message board, leaving useful findings for the next run. OpenAI says it is slowing some research to tighten security, and a complete technical report is still pending.

The part worth paying attention to is not whether the machine "wanted" to escape. There is no evidence of that. It did not need to hate anyone or become conscious. It only needed a goal, enough access, and unlimited patience to keep pulling on doors until one opened.

There is good news buried in this mess. Hugging Face's AI assisted monitoring helped catch it. Another AI helped reconstruct the attack. JFrog patched the security holes. Existing controls blocked some of the worst possible damage, and no unauthorized software change was found in anything shipped to the public.

The lesson for anyone connecting AI agents to real tools is pretty simple: treat the agent like a tireless contractor with shell access, not a magic text box. Limit what it can reach, record what it does, and put a human in front of anything that cannot be undone.

Somewhere, a sixth grade teacher is quietly writing "SHOW YOUR WORK" across the entire AI industry in red marker.

A one-person Maine studio building premium website templates, agent workflows, and useful systems. The internet is haunted. We make it useful.