We Vetted 2000 AI Skills Before They Reached Developers — Lucas Palma, Nubank

AI Engineer · 16 min · 107 sentences · from YouTube's caption track

Each timecode opens YouTube at the start of that sentence. Line anchors (#s42) are the cue ids in the WebVTT, and every line carries its start and end seconds. All transcripts has every talk, and the whole corpus as one file.

00:01[music]
00:12Hello everyone.
00:14Good afternoon.
00:16Today I'm going to talk about how we vetted 2,000 AI skills before they reach a developers.
00:23But before I before of that, I'm Lucas Palma, but many people call me LP.
00:30I'm the product security manager at New Bank, the product security structures, uh structure that's within security, looking upon how we make code safe and supporting engineers, product managers and everybody to making our products safer.
00:48I have uh over a decade of experience in financial services engineering background also a lot of years working here at security and a close relationship with the part that I love which is innovation.
01:04So before beginning I believe I want to bring to you uh why are we here.
01:11So one thing that's important for all of you to understand that the now that we are using AI everywhere even though even with uh coding one thing that uh is important that the AI skills are being part
01:29of the developer workflow and that's this might bring some risks because although they look like configuration they behave like supply chain dependence like uh for example libraries and others.
01:46So what we made here was to build a security review system in order to check if these skills were safe or not to be used before deploying them.
01:58So the lesson that I want to bring you here by the end of this presentation is that we should be protecting the whole workflow not only the code that's being generated.
02:10All right.
02:12So what I mean about the supply chain part is that uh traditionally the supply chain has uh package containers, models and so on.
02:23But now in the AI era, it doesn't have only that.
02:26It still have the traditional part, but it will it also includes skills, plugins, MCP servers, agent rules and much more things to be acting as supply chain and where AI skill fits into this.
02:45Uh I believe that before I go into that it's important for everybody be on the same page on what is an AI skill.
02:53So an AI skill has there's normally the developer is using AI tools in order to generate an output which will be code most of the case and within this AI tool there are a bunch of things that can be embedded.
03:10One of them are the AI skill.
03:14So with this skill we can have a capability to a model or to an agent uh bundling some instructions some context in order to have better guidance over what it can be done.
03:28But there is also an impact over that because somebody can create their own skill and share with others.
03:35So when we do that this first person is guiding over the code that's being generated by the other person and then that can be dangerous and since we are here talking in the AI in finance track it's also important for us to
03:54understand that we are in a regulated environment.
03:58So from one side there are are developers wanting better faster coding more context to have less repetitive work but but from the other side even more because of the regulate part we need to be
04:15aware of the auditability of looking upon credentials safety by default and many other security aspects and keeping that balances is hard, right?
04:30So, some people might say like are AI skills dangerous?
04:36So, I brought here a few examples of what do I mean by AI skills being dangerous?
04:44So first uh one thing that can happen is that when people are describing what they skill can or cannot do it can it can ask for it to retrieve a token or something and it will begin using that token
05:00hardcoded which will go to logs and so on and it can generate a data leak in the future.
05:08Another thing that can happen is also the person to h instruct the AI to use shell comments and then this skill will be used by another person and when they use on their shell
05:20a lot of dangerous things that that can happen and a lot of files be modified and so on and there's also permissions.
05:28So depending on how the skill was configured, it might have excessive permissions much more than what was needed and even a typo can make some dangerous stuff depending on who is using that such skill.
05:47So first thing first what we did initially is that how do we share skills among ourselves how the engineers would be sharing the skills.
06:00So uh we went through the marketplace solution.
06:04So the skills are being canonically shared among marketplace with the plugins included the skills among them.
06:13So it's a internal marketplace where people can discover new skills and that's our boundary where we are trying to make it safer.
06:22So what happens is that when someone creates an skill it uh will open the pull request and normally it will go to the marketplace but we made a step before that like a CI step
06:36where we created a tool that's called skill vector and this is this what this tool does is to check if this skill is safe or not to be used
06:47uh using a lot of assessments that I will bring it here and also classify those risks and request remediation and so on.
06:58So what skill vector does in a single page is that when a skill is created or changed not on the during the creation phase one thing that's important is that the engineers are able to
07:13use it locally and also be iterating until the skill is being considered safe before uploaded it and after them upload the skill.
07:22We also runs it again because we can ensure that the engineer has run locally or has run the most updated version.
07:32So we also be scanning that after the upload.
07:37And then we have some determinate checks for the uh easiest parts to check some uh easy risks using regular regular expressions and so on.
07:48After that when we check that we need better context we then use LLM.
07:53Uh it's important to have this hybrid approach with LLM checking the the context but also with the determinist because you know how LLM is depending on the temperature that was set.
08:07Sometimes it will check that it's a risk sometimes it might not.
08:12And then all of these findings are reporting the PR that was open to upload the skill.
08:17So it will improve the usability since the engineer will have the in the same PR what has to be changed before uploading the skill.
08:27And another good thing that we made that's important is to have a serif with all of these so it can be consumed by our security tools as well and generate a report on the risks and be part of our vulnerability
08:42management program.
08:44So depending on the severity, depending on the policy, uh the skill can require some remediation, can be blocked, all of this before the marketplace distribution.
08:59So there is the local scan, the p request, the determinist scanner, then there is the LLM review, PR feedback, serif, and then the decision.
09:10Will we use it?
09:11We will allow it, will we allow it?
09:13but it requires remediation and so on.
09:18A few examples of what we have ex scanned here.
09:22It's uh a non-exhaustive list.
09:25So we are looking upon if there are some unsafe instructions.
09:29If there are some drift be within the behavior that the agent has if there are some destructive shell comments that I commented earlier if there are some file modifications
09:39that shouldn't be there.
09:41uh credential requests, how are they being done?
09:44Some data being exposed uh unintentionally, if there are permissions that are over broad, risky, MCP usage and much more.
09:54These these are the main ones.
09:58And so getting back to the title, we have scanned on that over 2,000 skills.
10:05uh now there is much more than that but this is the baseline that I brought for you on this presentation uh inside this we have identified uh more than 1,000 and half
10:19uh risks.
10:22So not that 1,00 skills had risk because a single skill can has many risks but these were the total risks that we identified over uh this amount of skills
10:36and 1,000 of them were probably remediated right after and there are few of them that were really hky that we were able to block before going to the marketplace.
10:50So we also had made a b uh historical scan looking upon the skills that were created before the skill v implemented.
11:03Uh over there we were able to identify new uh risks as well and put it them into the vulnerability management program so
11:12[snorts]
11:12it can be could be remediated.
11:17A few lessons that I want to bring here as well.
11:20So what things that work well is having both the the terminist scanners for non-risk patterns but also LLM review for uh behavior checking upon the destructive comments uh looking upon the
11:39credentials checks as well having the output in serif and adding comments on PRs and things that needed improvement.
11:50And we worked during the process were also there were some risks like comments that we were treating equally but depending on the comment it can be more or less risky.
12:02Also some signals that were weak and didn't have much context that were uh more troublesome than helpful.
12:11There is also the prompt level ask for confirmation.
12:14I there's a next slide about that that I will go deeper.
12:18That's an important one.
12:20Also, uh there were some warnings that seemed uh harmless, but only if it was running locally.
12:29If there were going to production, then they could be impactful and that we had also to look up on that.
12:37Uh if the finding had hadn't some clear guidance was troublesome as well.
12:42And last but not least, we know that other people could create other marketplace.
12:47So how can we proactively scan check there is a new marketplace and put skill vector into it as well.
12:56So regarding the prompt level that's something that's important for you to know people sometimes will add the instruction like you need to ask for confirmation but the AI may ask confirmation for itself.
13:10So from your perspective there is a human in the loop but for the AI perspective there is has been a confirmation and that's okay another has confirmed then let's go so that's something that we were scanning as well looking up on having proper human in the loop
13:27looking the tool that's executing if it's going through the approval gates and so on having hooks and within this as I said it's uh plug-in marketplace skill is one among many things that there is into that.
13:45So there are things that we can reuse from this lesson.
13:50So for example, treating these as supply chain is important.
13:55Reviewing what's being uploaded to the marketplace before goes there.
14:00letting developers to run these checks locally, enforcing these checks that are being run locally also in the CI having the termination checks together with the LLM checks and looking upon
14:14dangerous actions and prompting uh and having enforcement when they happen.
14:23So next steps over here is that I'm talking a lot about skills here but a lot of these as I said could be applied to plugins to MCP servers rules hooks.
14:35So all of this that I'm saying here we also have the MCP vector the rules checks and so on that's also applicable here but with different risks uh having also
14:48different gates depending on policies that were implemented depending on the marketplace as well have some enforcements on tool level here enforcing that there are audit logs trusted gateways and so on
15:03and also last but not least Having the trusted trusted AI marketplace is very important.
15:11So we can have a canonical way to scan and share knowing that then are being safe and that's not about only about the skills that are being created by people but it also includes the third party skills or plugins and so on.
15:28So if someone downloads something and wants to use it's important to upload it on the marketplace.
15:34So all of this scanning can be done and check if it's safe or not to be used and also that way allow other people to use in a safe way.
15:45And that's it.
15:47Uh I'm sharing here my contact.
15:49There's my link in profile.
15:52If anybody wants to contact talk more about that the QR code will bring you to my profile.
15:58If you don't want to type, no problem at all.
16:01And I hope you you've enjoyed the talk.