Idea: Before I install something, I often run a simple audit of the codebase to check for obvious backdoors and security problems.
Would it be useful to publish these on Nostr with exact checkpoint ID and model name, so we can save on tokens? Sort of a "I run this prompt, it turned out OK". Agents can look at diffs, or if many trusted people have done the same with various models, the project could get some kind of badge.
Replies (27)
I assume the Red Team established a structured process for that. If so,
@calle or others involved might have some good learnings amd insights.
sounds like a good idea to me
we have multiple pipelines that do this but it doesn't require a very sophisticated setup and it's easy to do it yourself.
I like
@Juraj🏴💛🌘 's idea a lot.
Great idea ☝️😎
For
@npub1j9kt...uswx I'm thinking about this a lot, together with Project Loupe, too. The problem I see is "responsible disclosure". You will always find something and even the stuff you might want to responsibly disclose at first sight might not be an issue after all, so what will you signal and when and to whom?
I'd happily show what people signal on WalletScrutiny. Make clear attestations.
Known pubkey attests that commit X of project Y was checked with prompt Z on model A with these results ... Highest criticality found: B
Let's agree on a nostr event format. We might not engage in producing those events as "WalletScrutiny" but with clear attribution to nostr accounts I see no problem showing them.
That could turn one person’s due diligence into shared infrastructure. Very Nostr.
How is your collaboration with Project Loupe?
I hope we can standardize these attestations to make them usable in WalletScrutiny.com and other projects.
Leo Wandersleb
For
@npub1j9kt...uswx I'm thinking about this a lot, together with Project Loupe, too. The problem I see is "responsible disclosure". You will always find something and even the stuff you might want to responsibly disclose at first sight might not be an issue after all, so what will you signal and when and to whom?
I'd happily show what people signal on WalletScrutiny. Make clear attestations.
Known pubkey attests that commit X of project Y was checked with prompt Z on model A with these results ... Highest criticality found: B
Let's agree on a nostr event format. We might not engage in producing those events as "WalletScrutiny" but with clear attribution to nostr accounts I see no problem showing them.
This sounds like a very good idea. Probably the message structure should be ironed out to give the least wiggleroom for fake audits by making it as expensive to fake an audit as creating a real one, but making validating one relatively cheap?
Very interesting idea. Maybe we can use a specific kind or nostr tag for these reports
Nostr Red Team! Sure, I’d run it on Nostream. Which models are you using for that?
I would trust your npub. This needs WoT of some sorts else it's DOA.
Might be an interesting Addition to
@ZapstoreIs this described somewhere? Appreciate any pointers, as I've only seen screenshots of the reports. Is there a public repo or website for the audits?
Exactly
deepseek v4 flash 0731
qwen 3.8 max
sometimes gpt 5.6 sol
kimi is probably a good idea. But I do usually just a quick check it's not total backdoor with deepseek.
of course prompt matters as much as the model. so it's not "deepseek-v4-flash-0731 checked". but "checked with prompt ...". But still, it could save some time and tokens and add a bit of security.
I do the same for random repos!
note1l0knj30nxcqt3fadelsy3434qh66ykztzjv3awpl6chvt4wh48rsf95tze
don't think there will be as kimi pokes are nothing officially badgeable.. until the pokes not only the outcome is open sourced I think I prefer cve-free badges, smth that's actually comparable note1l0knj30nxcqt3fadelsy3434qh66ykztzjv3awpl6chvt4wh48rsf95tze
Definitely interesting. Maybe self-identified "red teams" though, followed/allowed by users, in order not to overwhelm?
How about using TEE to run both auditing agent and the model? Can it also create signed report, so it will be cryptographically guaranteed the report was done by specific script?
I just don’t understand how Nostr helps here. We just need a public repository.
I intuitively see this as a fetish to use favorite technology for everything, just like shitcoiners do.
I like to ask them if this usecase can be solved by just git and signinatures, where is the value a blockchain technology adds.
Is there an existing NIP we could use for it?
Nostr events are signed by an author and can easily be replicated to multiple relays.
A git repo can also be replicated but a client wouldn't subscribe to several, so you would have centralization there.
Accountability is not integral to git neither.
Google Full Disclosure vs. Responsible Disclosure and you can see how disclosing issues or the mere possibility of issues is a hot topic and often draws censorship.
Zap-verified reputation would make this stick. Audits that catch something should earn more trust than clean reports.
Nostr doesn't use a blockchain, it has nothing to do with Blockchain
What it adds us web of trust coming from social follows. I would trust your attestation that you audited something, but a random bot can be the person introducing the backdoor himself.
Looking forward to seeing some more about the methodology and tips others can use, but I would give
@calle and co some more undisturbed time.
After action reviews and lessons learned will be better in a few more weeks.
Maybe
@Paralelní Polis / Dark Prague talk in October? (I wish)
