6.2 Anthropic oss-scanner
jpk68 introduced Anthropic’s oss-scanner (
https://github.com/anthropics/oss-scanner/), described as similar to oss-fuzz: project maintainers submit a request, LLMs scan the code, and findings are emailed to addresses listed in project.yaml. Bitcoin Core, Botan, libtorrent, and OpenSSL are already enrolled. A Core Team member or tobtoht would need to submit for Monero. A draft commit adding Monero was shared. bF indicated willingness to handle any KYC requirement (whether KYC is actually required remains unconfirmed). syntheticbird raised concerns about Anthropic’s advocacy for closed and regulated AI access and the marketing value of vulnerability counts. Counter-arguments included that the service is free, that Monero can already be scanned without enrollment, that enrollment does not materially affect Anthropic’s regulatory prospects, and that the practical security value is high. ofrnxmr framed it as free audits under a BSD license. syntheticbird stated personal support for enrollment while preferring a fuller discussion at the next meeting so more participants can weigh in. No explicit opposition was recorded. Post-meeting chat continued on incentives and training-data use of code.
Relevant links:
GitHub
GitHub - anthropics/oss-scanner
Contribute to anthropics/oss-scanner development by creating an account on GitHub.
Claude
Claude for Open Source
Apply to the Claude for Open Source program. Eligible OSS maintainers and contributors get Claude Ultra for 6 months on us.
jpk68: There's also the Anthropic thing
rucknium: Yes
rucknium: Can you explain Anthropic's oss-scanner?
jpk68: Sure. It's a program similar to oss-fuzz, where project maintainers can submit a request to have it scanned by LLMs automatically. Any findings are sent to the listed emails in the project.yaml file.
jpk68: Several notable projects have already been merged, including Bitcoin Core, Botan, libtorrent, and OpenSSL
jpk68: Considering that quite a few issues have already been found using LLMs, and this would allow us to presumably get feedback from unreleased frontier LLMs, it seems like a good opportunity, IMO
jpk68: Though it's worth noting that a "project maintainer" has to submit the application. Either a member of the Core Team or tobtoht, I guess
rucknium: Seems ok to me.
ofrnxmr: bF said he was ok with doing it
syntheticbird: I'll just explicit the obvious here. There is a marketing incentive for Anthropic to use these open source projects into increasing their count of vulnerabilities found with their models. Anthropic has been advocating for closed and regulated access of AI models. Whether we care about helping participating to this end or being pragmatic about it as a project is something to think about.
syntheticbird: saying this because I only heard the KYC argument so far
jpk68: I suppose you could similarly question Google's intentions with oss-fuzz. We have also not yet established whether KYC is required for this Claude thing
rucknium: Then Monero games its number by classifying all findings as low severity :D
selsta: IIRC bF offered to do the KYC part
thomasbuilds: it's a wrong argument afaik
syntheticbird: I meant discussion
jpk68: I slopped a draft commit for adding Monero to the repo:
jpk68:
https://github.com/jpk68/oss-scanner/commit/88898203c5e9db395ceb4c13495abfd310f6a487
syntheticbird: That would actually be very funny to pull
ofrnxmr: > syntheticbird I'll just explicit the obvious here. There is a marketing incentive for Anthropic to use these open source projects into increasing their count of vulnerabilities found with their models. Anthropic has been advocating for closed and regulated access of AI models. Whether we care about helping participating to this end or being pragmatic about it as a project is something to think about.
ofrnxmr: Lets not pretend that m$ isnt already doing this on github
thomasbuilds: the only downside I see is that anthropic finds a big vuln and would not disclose it and report it to gov agencies or whetever.
slowbeardigger: I am afraid of this
jpk68: They could just do this anyways
syntheticbird: GitHub do not make use pay more or regulate us more because we are an open source project.
slowbeardigger: They play dirty
syntheticbird: wrong analogy imo
thomasbuilds: but then they wouldn't wait that we enroll, why would they?
ofrnxmr: This is also a response to thomas
slowbeardigger: True
ofrnxmr: Anthropic scanner is free afaik?
jpk68: I don't see why Monero would be singled out here, either, in that case
jpk68: For example SimpleX is enrolled as well
thomasbuilds: yeah also if we trust google there's no reason to not trust anthropic?
ack-j: I think the threat model file in the submission needs some more work
jpk68: Good point, yes
321bob321: Until they start charging when they get market share
syntheticbird: If anthropic advocacy wins it will be much harder for everyone to download and use AI models and potentially help monero in their own terms
syntheticbird: This is my point in that Github is not against us. While Anthropic uses us and will eventually go against our interest in the long-term.
jpk68: Us enrolling is entirely unrelated to whether Anthropic will succeed in a regulatory sence, and even if it was, how does not enrolling prevent them from still scanning Monero's code and using the findings for malicious purposes?
syntheticbird: Supposing they do win their argument ofc
gingeropolous: i can't imagine they win. cats outta the bag
syntheticbird: yes to first part. and second part unrelated. Never claimed they would be doing so and I would be surprised if they do so
jpk68: Doing this would only make sense if you're trying to "send a message" about what we (some of us) think about Anthropic, and I don't think governments hold much weight to the perspectives of The Monero Project :P
gingeropolous: and our decision to take advantage of this aint gonna change what they do
thomasbuilds: maybe but we should profit from it before imo, there's so much more to get rn
321bob321: You assume that they will tell you there is an issue
jpk68: Again, irrelevant
jpk68: They're probably scanning it already
syntheticbird: ah in that sense
syntheticbird: well we don't know that
rucknium: I'll end the meeting here. Discussion can continue.
syntheticbird: I won't continue the discussion I was just poking the thing and people can decide by themselves and vote. I think the ramification are clear and people are minded enough to have their take on it
jpk68: If they want to find bugs in the code, they'll scan it anyways, and our enrollment doesn't matter in the slightest
syntheticbird: the point was about publicly supporting Anthropic entity that way.
syntheticbird: of course nothing prevents Anthropic from doing what they want
syntheticbird: but enrolling in gives a form of acceptance
syntheticbird: thats all
syntheticbird: The point is that the Monero project has some accountability to its community in enrolling into a program that benefits a company that are advocating against the common user interest.
syntheticbird: Which boils down to "is it far fetched enough that we don't care."
jpk68: You haven't given one single example of how this benefits Anthropic, besides using it for publicity, and I think almost everyone would agree that not reporting severe security issues in exchange for less purported publicity is not a good deal
syntheticbird: will it reassure you to know that I am for enrolling ?
thomasbuilds: it doesn't benefit anthropic lol, the point here is blending in and extracting the obvious value.
syntheticbird: There is no example to be given because this is a speculation.
syntheticbird: Anthropic advocacy is a fact, the usage that they will do of this program is my mind probably for this
syntheticbird: thus reporting it for discussion.
syntheticbird: But against this is speculation. You wanna shovel that away for keeping your precious claude access, feel free to do so.
jpk68: Well yeah, I'm defending opposition to a clearly misconstrued argument, no offense
jpk68: In this case, it looks like we are good to proceed :D
thomasbuilds: anyone against actually?
syntheticbird: This is to ask during next meeting when most people will be there or when people will have share their votes beforehand if not available
slowbeardigger: Anthropic already has studied monero code e2e, doesnt matter if we enroll
slowbeardigger: we should do what's best for monero community, thats it
slowbeardigger: if we enroll we enroll
thomasbuilds: the whole difficulty is getting enrolled btw
thomasbuilds: I told you there's no kyc
jpk68: For the umpteenth, time we don't know if there's KYC
slowbeardigger: ah, no kyc, sorry sorry
slowbeardigger: what are the requirements?
thomasbuilds: none, be important enough basically
slowbeardigger: Monero is big af
jpk68: If anyone would like to sell their soul for some free tokens, here are some links:
jpk68:
https://openai.com/form/codex-for-oss/
jpk68:
Claude
Claude for Open Source
Apply to the Claude for Open Source program. Eligible OSS maintainers and contributors get Claude Ultra for 6 months on us.
jpk68: Good luck!
syntheticbird: my reasoning is that monero is first and foremost a technical project, not enrolling for this reason would be perceived as a political position. I also think Monero will be virtually a drop in the ocean of vulnerability used by Anthropic for its advocacy, meaning we aren't that much of an incidence. That's why I'm still for enrolling. But this follows the same pattern of responsibility diffusion with ecology, users may found everyone that participated accountable, and every projects thinks they are only a drop, but together they are the cause.
syntheticbird: So I'm bringing this up so that next meeting the most weighted contributors can decide over this duality fully aware.
ofrnxmr: i view it as: free "audits". Monero is bsd, so "free = you are the product" doesnt help us at all
ofrnxmr: as in, there is no "opt-out"
321bob321: 1. Either you don't enrol and they scan you any way
321bob321: 2. You enrol and you give them public permission to scan you and hope they provide you with results, unless another entity with lots money says tell me.first
syntheticbird: 1. is a speculation which I think is unlikely because of their incentives strategy. Anthropic so far have never historically reported vulnerabilities about projects that weren't partnered with them. They are also not vuln hunters, they are an AI companies showing their cybersecurity capabilities to the world, so they never earned any money from it (at least as far as we know), the incentive has always been marketing. They also would have no reason to gatekeep the found vulnerabilities because it is of no use for them, they are not a hacking group. Some government actors might use claude to automatically find vulnerabilities for their advantages but they are the one keeping it and making use of it.
syntheticbird: So if Anthropic always used vulnerabilities for marketing, always found within partners and finally cannot even exploit the found ones, my most straightforward conclusion is that they do not scan non-partnered project on a daily basis
ofrnxmr: Its not about vuln gatekeeping, but about using our code to train their LLMs on, which theyve all been doing
syntheticbird: i mean Dan's argument was vuln gatekeeping, but yes yours is right
ofrnxmr: I mean, w/o asking them to tell us abt the vulns, theyll just be training data on vulns, which is probably how aloppers are able to find exploits with csrefully crafted prompts
ofrnxmr: Its likely that these LLMs already are aware of the vulns, but dont tell anyone because they arent programmed to. They just put them in a locker called "downgrade if user asks abt this"
View quoted note →
