redthing1

TEE stuff @zellic_io / i like TEEs, ORAM, and applied crypto

[redthing1 @rt1rz]

Compromising Signal's Contact Discovery Enclave
https://t.co/9m7nsTtFhq

[V12 @v12sec]

Signal's Contact Discovery automatically sends your contact list information to an SGX enclave in the cloud.

V12 broke into that enclave and leaked the key, allowing the server host to decrypt everything.

Two separate critical bugs: arbitrary read and RCE.

Here's how. 🧵

[s1r1us @S1r1u5_]

optimistically, lets assume cyber is defense dominant and exploitable vulnerabilities are finite and exhaustible.

then, give the fucking models to as many good security researchers as possible and let them exhaust the attack surface.

pessimistically, i think cyber is offense dominant. the sin of complexity, vibe coding making it worse, and increasingly capable models potentially opening up entirely new vulnerability space, and you need to secure everything.

then, your best bet is formal verification and secure by design, but thats a fever dream and aint happening anytime soon.

so, again, give the fucking models to as many good security researchers as possible and let them find and eliminate as many attack paths as they can.

either way, walled gardens are counter productive.

[s1r1us @S1r1u5_]

don't trust a dev to reason about complex software to keep it vulnerability free. this might sound a bit weird, but hear me out.

up until now the playground was mostly level, and everyone operated with roughly the same cognitive capacity. security researchers had an edge because they were incentivized to reason harder and longer than developers, but the gap usually wasn't enormous, there were only a few of them, and a lot of software never got attention.

now that we have entities capable of reasoning beyond that capacity and dont get tired, all our old complex software will keep on giving.

the problem with JITs, image parsers, emulators, and similar software is that is their complexity blows up it takes far more cognitive effort to understand every possible state than the software took to build which only needs few things to know

you can build them with just-in-time cognition. reason about the piece in front of you, make it work, move on. but understanding every state the final composition can reach requires sustained reasoning across the entire system. no designer or developer does it. it's painful, expensive, and unnecessary for producing software that works.

this is fine as long as the program never processes untrusted input, or the world has no adversaries. the moment either changes, there is an incentive for a group, or even a single individual, to do the sustained reasoning, understand the composition, find vulnerabilities, and exploit them.

with every better model, a good amount of previously inaccessible vulnerability space opens up. look at the number of issues chrome fixed after mythos release.

this is interesting because anyone with a better reasoning machine than anyone has an advantage over anyone who doesn't. look at astra popping v8 bugs that gpt 5.6 sol and mythos missed.

[Matthew Green @matthew_d_green]

More TEE attacks, this time on Signal’s contact discovery enclaves. https://t.co/4tNRDSUbXT

[redthing1 @rt1rz]

@saxenism @trailofbits imo, (1) no excuse to write TEE code in a non memory safe language (2) AI will commoditize software side channels

[redthing1 @rt1rz]

@MysteryHacker1 @BitchatMe_ @jack as far as i remember, you can’t disable sending contacts to CDSI either, it’s unconditional if the app has the contacts permission.

[redthing1 @rt1rz]

@lyq_sqsp @elonmusk brilliant and impressive work!

[redthing1 @rt1rz]

@MysteryHacker1 @lyq_sqsp it exists so we can cite it as a practical use of oblivious computation in production in our paper intros :)

[redthing1 @rt1rz]

@lyq_sqsp thanks! yeah… had the same thought. sgx is so well studied that all the attack primitives are known, but as usual there’s a big gap in academia -> practice, and fuzzing for these bugs is tricky.
i linked my USENIX ORAM paper though :)

[redthing1 @rt1rz]

@yanjulang i’m very glad it worked :D

[redthing1 @rt1rz]

@saxenism @trailofbits We have more coming :)

[redthing1 @rt1rz]

@saxenism dsv4 flash 0731 is surprisingly capable

[redthing1 @rt1rz]

@YadKonrad Heavily driven by me, I do research on ORAM/SGX, the models lack domain knowledge here.
Agents are awesome for digging at cracks you find, rapidly turning a bug into a full-fledged chain

[redthing1 @rt1rz]

@YadKonrad read the writeup
https://t.co/m9A30cNHlI

[redthing1 @rt1rz]

@YadKonrad it's real! research done by me. repro'd on actual hardware

[V12 @v12sec]

Signal's Contact Discovery automatically sends your contact list information to an SGX enclave in the cloud.

V12 broke into that enclave and leaked the key, allowing the server host to decrypt everything.

Two separate critical bugs: arbitrary read and RCE.

Here's how. 🧵

[BinBin @binsquares]

Project is https://t.co/tGV3kZcdqO

[redthing1 @rt1rz]

a new lldb tui i'm working on. source on my github.

[redthing1 @rt1rz]

@thsottiaux i'm especially impressed because codex is the only model that can competently handle complex things like LLVM passes. other models like claude/gemini fall over themselves and hallucinate blindly

[redthing1 @rt1rz]

@momo5502 impressive! can't wait to try it