Enroll vaultwarden in anthropics/oss-scanner

[gaby]

@dani-garcia Anthropic now offers free security scans for OSS projects, you can enroll by sending a pull request here: https://github.com/anthropics/oss-scanner

They can send you the report via email or as github security advisories.

[GN998]

[GN998]

Fair enough, up to the maintainers either way.

[BlackDex]

However, Anthropic's terms are quite unfavorable to developers and will impose additional burdens on them.

@GN998 may I ask what is unfavorable and which additional burdens you think are going to be imposed?

[GN998]

@BlackDex

My main concerns come down to a few points:

• Triage fatigue from LLM hallucinations: LLMs tend to treat ordinary bugs as security issues—like self-inflicted issues that only impact the user themselves rather than cross-user (User A affecting User B) vulnerabilities. Filtering through a high volume of these reports will be a huge drain on maintainer time.
• Upstream communication blocks: Under the terms, reports are confidential and can't be shared outside the project's maintainers. If a vulnerability actually lies upstream, we can't just pass the report along to them.
• Asymmetric liability for individuals: Anthropic caps its liability at $1,000, while individual maintainers have to shoulder the compliance burden. If a maintainer's email or GitHub account is compromised and a "confidential report" leaks, the maintainer is technically on the hook under the terms.
• Training on feedback: Any feedback we give on these reports can be used to train their models. To me, this feels like we're acting as free data labelers for their scanner while bearing enterprise-grade confidentiality risks. I'd suggest looking into whether OpenAI or Anthropic offer more OSS-friendly programs where interaction data is strictly excluded from training.

Of course, this is just my two cents. If the maintainers feel the program is worth it overall, I completely respect that decision.

[BlackDex]

@BlackDex

My main concerns come down to a few points:

• Triage fatigue from LLM hallucinations: LLMs tend to treat ordinary bugs as security issues—like self-inflicted issues that only impact the user themselves rather than cross-user (User A affecting User B) vulnerabilities. Filtering through a high volume of these reports will be a huge drain on maintainer time.
• Upstream communication blocks: Under the terms, reports are confidential and can't be shared outside the project's maintainers. If a vulnerability actually lies upstream, we can't just pass the report along to them.

I'm not sure of you have seen the PR explosion, but most of that is LLM already. And we also already get enough LLM generated/found security reports from different people, and also multiple the same, so it's already here.

• Asymmetric liability for individuals: Anthropic caps its liability at $1,000, while individual maintainers have to shoulder the compliance burden. If a maintainer's email or GitHub account is compromised and a "confidential report" leaks, the maintainer is technically on the hook under the terms.

I have signed no contract to pay or be responsible for anything. It's open source software, everybody can use LLM's to find security issues and abuse them if they have bad intensions.

If my email or GitHub account is hacked, i probably have bigger issues.

• Training on feedback: Any feedback we give on these reports can be used to train their models. To me, this feels like we're acting as free data labelers for their scanner while bearing enterprise-grade confidentiality risks. I'd suggest looking into whether OpenAI or Anthropic offer more OSS-friendly programs where interaction data is strictly excluded from training.

Both companies already have all our code, either since the maintainers have used an LLM or other security researcher's have run out codebase through several of those companies tools. Also, it's open source, so that kinda is the whole point.

Of course, this is just my two cents. If the maintainers feel the program is worth it overall, I completely respect that decision.

[GN998]

Totally get where you're coming from! Dealing with the current wave of LLM PRs and duplicate reports is already a lot, and maintaining Vaultwarden is no easy task.
Thanks again for sharing your thoughts!