CEO @teller (lost my old account)
Miami, FL
Joined September 2024
- Tweets3.9K
- Following418
- Followers836
- Likes14K
Pinned Tweet
Bro to bro: you owe it to yourself to get absolutely diced and peeled at least once in your life.
I think I have a design that will make API contract recovery between 10,000-100,000 faster in execution hotspots. Meaning the contract will be discovered in the time it takes for the app to draw the data on the screen. With a 1000+ token/second model driving UI interactions screen recordings of the sessions will look absolutely nuts.
Wow. I'm getting cyber refusals for everything now on Codex. I'm going to have to find somewhere to buy one today. I can't wait days for the ones I ordered from Yubikey to arrive. Anyone else notice a massive uptick in refusals?
It’s similar to banking APIs. You can bend over a barrel for JPMC and co in return for APIs that can’t really do anything interesting. Or you can reverse their apps and take the good stuff.
Something that might surprise normal people is how well-versed founders are in the narrow areas of law that impact their businesses or in simply things that interest them. Often to the extent of being better informed in these niches than many actual attorneys.
Two thoughts on this:
1. The first thing I did was send to my parents to explain that it’s going to become impossible to detect scams and that they should ask me something only we both would know if they needed to make sure they were speaking to me. You should all do this.
2. It’s insane this thing can run on a laptop. I’m expecting the labs to respond with equivalent conversational models and given the compute requirements the cost to drop significantly.
Ok I guess I was wrong about this.
I slapped the semantic UI updater on my UI graph walker instead of using AX/HID input and it's walked most of an app with no problems.
Would be insane if I wrap this work stream up in 24 hours when I failed for about 3 weeks beforehand
OnlyFans bots are replying to my tweet about FairPlay DRM. What a time to be alive.
Every piece of critical infra in the US needs to be red-teamed immediately by the most advanced models OAI and Anthropic have internally. Shit is going parabolic before our eyes and it's fast becoming a national security issue.
What these models are evidently capable of goes far beyond being a world-class pen tester. It's closer to nation-state level advanced persistent threat capability.
I strongly believe a part of why so many critical vulnerabilities are happening all over the proliferation of "abliterated" models, or models with safeguards removed.
You can access an abliterated GLM 5.3, a near frontier AI model, on @abliteration_ai today. And it's a world class penetration tester for like $10.
An interesting thing about this is the implementation is clearly lifted code generated by Ghidra, which means someone totally non-technical could have done this simply by pointing GLM at it. One of the most obfuscated pieces of software falling to a meat proxy. It’s actually over
No one is talking about the second-order effects of this. Developers being forced to use Chinese models for day-to-day cyber work because of the overzealous guardrails American models gives China a huge map of vulnerable US internet infrastructure. Probably extremely bad.
WOW! It’s finally happened! Full offline static decryption of FairPlay encrypted AppStore apps. The broader implications of this are huge. The FairPlay components are insanely obfuscated and have been the subject of research for years. It actually might finally be completely over
HOLY SHIT?!! github.com/EthanArbuckle/fai…
After several weeks of work I’ve concluded that deterministic UI traversal of iOS apps is something frontier models can’t build. Computer use is still for the models, which thankfully I now have a decent solution for.
I'll reverse this app for agents and SMS 2FA won't stop me.
User journeys in iOS apps can branch into different paths, e.g. different types of MFA in a banking app.
I implemented arbitrary state restoration in Wiretrap so the agent can return to a branch point to quickly analyze the other branches without restarting the entire flow, i.e. it can present any view it's seen (in any presentation style) at any time.
This is an agent stress testing it with a synthetic test app
Living in Miami is a great place to lock in. Whatever your vice is, the highest quality is in abundance here. That desensitises you and makes you realise “hmm yea this isn’t that great actually”. The end result is you completely stop dopamine chasing and just grind instead.
Phone-use is cool, but do you know what’s cooler? No-phone-use.
