
product engineering @bluethroat_labs ✮ ⋆ ˚。𓅨⋆。°✩ (๑ᵔ⤙ᵔ๑) building a desktop ethereum wallet ദ്ദി(˵ •̀ ᴗ - ˵ ) ✧
Joined July 2024
- Tweets380
- Following532
- Followers194
- Likes10.6K
Pinned Tweet
I built a desktop wallet to replace now unmaintained Rabby and Frame (≧◡≦)
(づ ̄ ³ ̄)づ links are in reply/profile
i'm in my mindful era: not trusting AI output, tracking minor changes and trying to deeply understand the protocol i'm developing
any tips for getting back to mindless prompting? my brain is cooked
1
1
59
wehi.crops retweeted
Lighter says the drained accounts used unofficial signing code that reused nonces.
That's all we KNOW.
My guess, pure speculation:
+ Someone asks a coding agent for a Rust trading bot.
+ The agent ports the signer instead of using an official SDK.
+ It generates ONE nonce at startup and reuses it for every order.
+ Every test passes. Every order fills.
+ Every signature quietly leaks the key.
Cheap fix every protocol should ship:
+ LLM-specific instructions on your docs and repos.
+ An llms.txt or AGENTS.md that tells any agent touching your infra to use the official SDK and NEVER roll its own signer.
Your users' agents are reading your docs.
Write for them too.
so my feed now is just @fede_intern arguing about new UBA program with VERY toxic people
i decided to take another look at it and maybe provide some adequate feedback
for context, i'm a math major dropout and had been teaching high school math for ~4 years, now work as an engineer
so, the questions:
1. if AI is allowed, how do you make sure students get the deep understanding of, say, calculus?
i'm not familiar with UBA traditions but oral exam may turn very dramatic if AI is allowed (students won't even understand why they failed). i suggest having grades for every seminar, this will keep students active and promote deep understanding.
2. how do you make sure that history, politics, etc courses are built exactly for complex systems?
it's a principal-agent problem. for a regular teacher it's just easier to teach a default history course but the program requires deep connection with other courses.
3. AI can make education worse. few examples:
not enough programming imo. it's a skill that people used to learn for years (assuming good level at algorithms), not one term. with AI students won't notice the impact of small bugs. are you betting on english-as-last-coding-language?
there's a lot of math. however, in my experience, with AI you can just slide through all the theorems. what you need is tinkering around, modifying necessary and sufficient conditions of theorems.
so this, 3rd question, is "how do you protect students from cons of AI education?"
4. no flexibility? some students may want to go deeper into ML/finance/biology
i understand that it's a new and very opinionated program but do you plan to add courses of choice? or are workshops flexible enough?
5. is it only for Argentinian citizens? are courses in Spanish or English?
i'm a foreigner and despite the fact that i might not be the best fit for the program i'm really excited about it
well, these turned out to be just clarifying questions. i love the program so far ⊹ ࣪ ˖꒰ঌ ໒꒱.⋆˚࿔
6
183
wehi.crops retweeted
Replying to @devanshmehta
This is definitely NOT going to be as straightforward, but in general, this is a good direction.
This will definitely help you catch implementation bugs, assuming the “Convert to lean with AI” is actually credible.
Notice that DeepMind didn’t do this step at all.
They proved properties directly on the compiled ARM binary of libsignal using a formal semantics of the machine code, not on an AI-translated Lean version of the source.
But in my red-teaming experience, I would say you require a quite different set of specs to break systems.
The proofs only show the code matches the spec; the breaks thrive in what the specs could not fully express. Even DeepMind kept humans in the loop for writing the specs :)
wehi.crops retweeted
"We use TEEs" is not a security guarantee.
Two teams can both run TEEs and offer wildly different protection. The details are where it breaks.
4 things we check every time:
1. Physical access changes everything
If your adversary can touch the machine, we treat the TEE as compromised. Published attacks like TEE.fail are why. Whoever hosts your hardware is in your threat model.
2. Attestation doesn't mean your code is safe
A TEE will faithfully run vulnerable code. Bug in the inference engine or a dependency? Data can leak without anyone breaking the hardware isolation.
The code inside the boundary needs review too.
3. Attestation is evidence. Your verifier makes the call.
Authenticate the report. Validate the collateral. Check identity and measurements. Enforce a real policy on platform security status. Then bind that decision to the endpoint actually receiving your data.
Importing an attestation library does none of this for you.
4. Is the audited code the code that's running?
Open source is a start. You still need a verifiable link from source → build → attested measurement. Reproducible builds get you there. Without that, "audited" and "attested" can be describing two different things.
Real example from our Phala/dstack research:
CPU evidence was cryptographically verified. GPU evidence was fetched, parsed as JSON, and shown under "TEE Hardware Verified" with no GPU crypto checks.
The GPU claim had no cryptographic backing. That path has since been fixed.
bluethroatlabs.com/reveries/…
Inference providers are under attack. Securing the hardware boundary is one layer of securing the service.
More real failures and practical checks in our TEE Security Handbook:
tee-security-handbook.blueth…
nitter.gq/ilblackdragon/status/2…
Inference providers are under attack.
AI inference is now playing a critical role and is a honey pot of data and ability to inject actions back into users devices via agents for malicious actors.
NEAR AI thwarted a series of attacks on its systems and is back to normal. NEAR Protocol, Intents and other infra wasn’t affected.
TEE architecture is part of the answer to making sure that from outside even inference provider doesn’t have an access. User should verify exactly what happens with their data and prompts.
NEAR AI code is open and inspected by many partners when deciding to leverage the stack to validate exactly what runs inside TEE and ensuring privacy.
In this case there was also a vulnerability within SGLang, open source inference engine that runs inside an enclave. Even though this had a minimum disruption it still was a very
Going forward we are going increase security rigor of dependencies and also improve a “turnstile” to detect when anything out of normal runs inside TEE.
Proactive security against external attackers has became a must. We have been using Abliterated models and red teaming harness to continuously monitor and detect issues. We are also expanding scope and compute invested into this.
so, how was devcon for me?
+ met a lot of people i only knew online (s/o @0xMattmatt @apoorveth, not even mentioning Bluethroat Labs)
+ met a few people from EF, turned out they are just normal people, not some kind of wizards
+ all the talks i attended were great
+ because of volunteering at security hub i attended even more talks
+ saw Vitalik... im a fanboy i know
+ surprisingly there were LOTS OF people from my country
+ made new friends!
+- i drank too much....
- food was too spicy for me, so i was eating plain rice all the time
- traffic is crazy, because of it i mostly stayed in BKC
and the main concern:
- i have a feeling that i won't be able to catch up with all the people i met because there are a lot of you guys and im an introvert
feeling happy
yo i've accidentally closed my dms but now they are open
so feel free to contact me on x as well
3
187
my trip to Devcon was great
but it wouldn't even near as good if not the @bluethroat_labs team
there are not enough words to tell how thankful i am
Lisbon is next (actually after Vietnam and Mongolia xd)
Looking forward to the same photo in Lisbon next.
We are always looking for exceptional talent to join us over at @bluethroat_labs, so we can have a bigger team XD
4
11
412
wehi.crops retweeted
Looking forward to the same photo in Lisbon next.
We are always looking for exceptional talent to join us over at @bluethroat_labs, so we can have a bigger team XD
due to *some disturbances* in the flight schedule i'm not attending this :(
so it's a good chance for you to gossip! send dms to @saxenism
2
4
270
btw you can use BluePaper with agents using bluethroat-cli
bluepaper.bluethroatlabs.com…
Congratulations, you got the job.
Now, do you open the attached PDF to sign the joining letter on your personal system?
BluePaper, our next free public good, renders the document in a disposable sandbox and returns a PDF rebuilt from pixels.
Live: bluepaper.bluethroatlabs.com
when i got a job at Bluethroat Labs it felt too good to be true so i was literally panicking – am i being scammed?? should i open the contract? how do i sign it??
now there's a solution for securely opening PDFs
Congratulations, you got the job.
Now, do you open the attached PDF to sign the joining letter on your personal system?
BluePaper, our next free public good, renders the document in a disposable sandbox and returns a PDF rebuilt from pixels.
Live: bluepaper.bluethroatlabs.com
wehi.crops retweeted
Congratulations, you got the job.
Now, do you open the attached PDF to sign the joining letter on your personal system?
BluePaper, our next free public good, renders the document in a disposable sandbox and returns a PDF rebuilt from pixels.
Live: bluepaper.bluethroatlabs.com












