It's /ʲustɑs/ • security research • machine psychology • privacy • FOSS
🌐 🇺🇦 🇦🇷
Joined January 2021
- Tweets2.4K
- Following745
- Followers1.3K
- Likes14.4K
Pinned Tweet
I climbed a few more places on the @ethereumfndn bug bounty leaderboard! A bit late to share, but still happy happy
+rep to the triage team, I see in real time how they've been improving the process
😬
Honestly, it's beyond me how there are still such things as "blind signing" (wtf does that even mean) and 3s long tap confirmations on each one of the 58 pieces of calldata any modern dex swap requires (with expiration!)
Is there some fundamental problem in allowing to scroll the calldata string on a 3-4 inch display?
Well, it now seems like the whole Internet was compromised in the last few days. Were there precedents of this magnitude before? Not sure, I must be too young to tell
I can tell, though, that LLMs and AI are hardly responsible if you want to blame them. The code made years ago had cancer inside that went undetected because we had no better diagnosis
I feel relief, if anything, because the demons the labs have been scaring us with for the last year are now real. This is no longer an uncertainty. We are out of the denial phase, and we can ask "What do we do now?" to step into acceptance
We'll iterate, review architecture and norms as the old ones are falling apart. Taking losses, but learning. And we'll build new ones as we always do when the world gets changed
Well, it now seems like the whole Internet was compromised in the last few days. Were there precedents of this magnitude before? Not sure, I must be too young to tell
I can tell, though, that LLMs and AI are hardly responsible if you want to blame them. The code made years ago had cancer inside that went undetected because we had no better diagnosis
I feel relief, if anything, because the demons the labs have been scaring us with for the last year are now real. This is no longer an uncertainty. We are out of the denial phase, and we can ask "What do we do now?" to step into acceptance
We'll iterate, review architecture and norms as the old ones are falling apart. Taking losses, but learning. And we'll build new ones as we always do when the world gets changed
I hope Ledger will update the hardware integrity page because the pictures there don't include all/latest revisions of the devices, which makes the comparison more difficult than it needs to be
Hi @wellington0x, to open a signer, here's what you should know first.
Opening the casing doesn't automatically void your warranty. A warranty claim only becomes invalid if the device is damaged as a result.
Signers aren't designed for routine opening, so be extremely careful, and use the correct tools. We offer support articles specific to each Ledger signer covering the main precautions. Each article also includes images of the official internal layout, including key components: support.ledger.com/article/4…
If you have lost funds - please ensure that before you open your device you have logged the issue with your local law enforcement and they have confirmed you can open the device.
If you need help or have questions, please contact us directly: support.ledger.com/contact-u…
Would be kinda cool to assemble your new hardware wallet by hand? Who'll be the first one to market like this?
Ledger Nano X can weigh anywhere from 34 to 36g, according to the official specs and past community measurements
Other models also vary enough in weight that weighing a device isn't a reliable check on its own. I guess it would also be possible to shave some weight off a device
Posts like this make me realize how far down the rabbit hole I am... And I don't even fine-tune or benchmark models at scale, so I must be mid in the grand scheme of things
Just to avoid some confusion: auto-compact does summarize. When it runs, your whole conversation is replaced by a short summary. It doesn't keep the last 1M tokens around.
The ~967K is only when it runs on 1M models. Compacting itself is just one request that reads about as many tokens as the message you sent right before it, and it's mostly cache reads if your cache is still warm.
What @rohit3a found is that waiting until 967K means every message before that point uses a lot of tokens, because each one includes your whole conversation so far. Most of it is read from the cache, which is much cheaper (so this is usually fine as long as your cache is warm), but it still counts toward your usage and adds up as your conversation grows.
/autocompact 400k makes it summarize at 400K instead, so your messages never get that big.
The latest changes to Anthropic's CVP have made Opus 5.5 a useful addition to my usual workflows. GPT's usual "skepticism" and Claude's "wit" balance each other out quite well (a few stronger words come to mind as I write this...)
If you want to set up long-running tasks and supervision in Codex, take a look at my toolkit, ferrumctl. I use it daily and update it regularly
It has now also proved useful for letting Claude communicate with Codex across harnesses. Unfortunately, not the other way around yet, but you can use local transcripts. Getting into Claude's internals is harder, but I'm working on it (I have a prototype that exposes the SDK as a CLI)
github.com/ustas-eth/ferrumc…
OMFG, now it seems you've got to NOT USE the daybreak flag in you requests to avoid triggering the safeguards.
What the actual fuck.
Thanks to the other unfortunates in the topic for sharing the tip
community.openai.com/t/daybr…
I just want to do my job in peace.
Regards
Would be so good to have any kind of communications on the matter from @OpenAI @thsottiaux
The fork is updated to 0.161.0+cyber.1
There are some improvements upstream with the opt-in features.cli_daybreak option to include the API flag, but AFAIU it still doesn't work for automated turns like the goal ones. Hence, the fork is still required to provide the global and per-model defaults
Works like a charm :)
Pushed a few updates to fix goals deblocking. There's a small edge case where the model can mark its goal blocked and get the capacity error right after that, while still working. The goal will be resumed in this case.
Unfortunately, I don't see any ways to catch it without stacking complexity or changing the Codex code, so I left a parameter to disable goal changes.
Enjoy
ALT the chat was resumed automatically on a capacity error
github.com/ustas-eth/codex-r…
This runner will resume all your threads automatically on capacity errors, with timed retries, no user messages, and no other unrelated state changes
Note: I'm still testing it myself, if you encounter any problems, please report on the GH or here in comments
Good direction, unfortunately Red is only available to API and enterprise users. The existing CVP authorizations, it seems, are Defense by default
We’re expanding our Cyber Verification Program to give security professionals broader access to our most capable models.
Through this program, verified security professionals can access Claude Mythos 5.1, Opus 5.5, and Sonnet 5.5 with safeguards designed for defensive work.
We’re also opening up new tiers to allow for authorized offensive work, like penetration testing and red-teaming.
anthropic.com/news/cyber-ver…
