Cursor for Security Teams
Learn how security-minded teams use Cursor to reduce bugs shipped, prevent vulnerabilities, and raise the overall quality bar, without slowing engineering down.
Terms in this video
Transcript
So, first of all, thank you everybody for coming. Uh, my goal with this is to really just show off how I've been using cursor and AI to do really cool security work. Um, if you are deep into AI and have been using things, some of what I'm showing here might be obvious, but I wanted to just level set. Uh, I think there's very very powerful tools today and I want to make sure that everybody at least sees what I see and I'm happy to have this be the start of a
conversation. Um, so this is me. Uh this is what we're going to be covering today. So very very quick. The the slides are going to end briefly in about five minutes after I explain like what this is. But most of this is going to be showand tell and Q&A. What I want to do is weave in basically after I I get all of this introduced. I'm going to show me using cursor last night. This is just a live screen recording that I did of me using cursor to do the kind of security
work I would do in a hypothetical scenario. So, I'm going to let the video run um as it is, and some of that will be boring. And so, what I'd like to do is weave in your questions as we go, and Emily has offered to be a moderator for this for me. And then finally, at the end, we're going to leave some time for just strict Q&A about any of this kind of stuff. Um, we can have separate sessions about how does uh the security pro program at Cursor operate, like
product security features. I'd really like to leave this one about how do you use cursor to do security work. Okay. So with that, um I guess like very very briefly, I'm Travis. I joined in June. My background is all security. Prior to this, I was running a company. Um my obsession really in the security space is we need to be doing security engineering. I think AI is this like wonderful tool where it handles enough of the engineering stuff that we'd want
to do and it's really a great partner for security. That being said, um I have discovered like my general tenants about how I operate with AI and I wanted to level set those here and you're going to see this in the in the demos that I show. So models need context. Uh that's and this is like very similar to if you've ever managed people. Uh if you want work done that you're not doing yourself, it's very important that you be clear enough about what that work is
and what good looks like and what you do and don't want to see. And models need the same thing. If you're very vague, they're not going to get it. they might go off in the wrong direction. You're going to get frustrated. Um, but you know, for anybody that was around in the early days and you had to do prompt engineering, you know, spend three hours thinking about exactly what to tell the model, you don't have to do that anymore. Um, so you'll see that today.
Um, seems obvious. LMS might be wrong. Um, you should just always assume that like the model might be lying to you or just fabricating stuff like that's the way they operate, but that doesn't mean it's useless. Like there are techniques that you use where you're always somewhat suspicious of the model. I think security people are uniquely good at this. Like we have suspicion of everything. So being suspicious of the model will come naturally to us. And it
doesn't mean that it's useless. It just means that you you build workflows around assuming that any individual fact you get could be wrong. And so number three, you should be asking for proof. So if the model says something and you're like, I don't know, I have doubts, you can always just push it for detail. Show me the code. Show me the snippet. Show me what whatever. Like make me an architecture diagram. Ask questions about the architecture diagram. Number four, no dumb questions. So I have been using cursor like one of
the biggest use cases I have is to just ramp into something and learn quickly about it. Um this is something that we always have to do in security, right? We we have like very horizontal coverage area. We're doing security for parts of the product or integrating security in systems that we haven't dealt with before. And models are wonderful at ramping in and teaching you about something. So there's no dumb questions with with agents. You can just ask until
you understand it. And then agents as you're going to see in demos are just extremely fast at search. So you should use that. Okay. So the scenario uh some developer you're you're doing security at a company some developer comes to you uh they really want to try OpenClaw. They really want it connected to their corporate email um and they want to text with uh OpenClaw from their phone. And your job is to give, you know, thumbs up, thumbs down, like is this cool to do? Um, I picked this because it's fun
and actually like wanted an excuse for myself to look at OpenClaw last night. And also it's timely. Like there's a lot of interest in this and I think we're all wondering to some extent like is this safe? It seems very cool, but the the security concerns are real concerns. So we're going to dig into that a little bit here. And so just the setup, you just clone Open Claw like I did last night. You fire up cursor. Okay, so now that's it. That's all the
context. Um, now we're in the demos bit. So, first thing, this is just like where I start. Uh, I just I'll I'll pause here so you can see the prompt. So, this is like my level of prompting. So, I gave it literally the pointer to the GitHub repository. And what I want to know is like how old is the repo? So, this is like a general security smell thing. Is it brand new? Um, who are the contributors to it? Tell me the top 10 and like number of contributions, how
long they've been with it and who they are. What are the goals of the project? Like I actually have heard stuff about open claw but I'd like to just get fresh eyes on like what exactly is this thing? Um basic hygiene. Is there security policy? What's CI like? Is there a well definfined development process? And then the top level architecture. Okay. So it's going to crank on this thing for a little bit. And what I like about cursor is that it's very very f like I'm not
here to sell stuff but like I like cursor and I use cursor. I could use other tools if I wanted. There's no mandate at the company. Actually I'll pause here. There's no mandate at the company. I have to use it. I use it because it's fast. Okay. So, why is it prompting me here? It's prompting me here because I'm paranoid. Um, so my GitHub user is admin and I do not want any state mutating um things to run without my approval. So, in the case of GitHub, this command, it's going to run with admin and we have various settings
that you can use like I'm running this in sandbox, but when it's coming to running this command, I want to approve it. And so, that's what it's doing here. And what I will do is just read this. Um, agent generating this command is very fast. Even if you are like a GitHub CLI expert, you would not be able to write it as fast as agent. So me spending one second to read it and make sure that it's sane is completely fine with me. Okay, so we're going to keep
going. I'm going to run it. I'm going to run this one. I'm going to run this one. Run this one. That's all fine. I am scanning all of these and making sure that it's sane. And now it's going to continue doing a whole bunch of research. So with that, do we have any questions? so far about any of this. I will pause at the next interesting point, but we can take probably a question or two. >> Yeah, I guess what is your ideal type of sandbox to run agents? Um, not just open claw, but just coding agents in general.
>> So, we in cursor we have a sandbox. Um, we I think that we've implemented that very well. But in general, like what you want is some kind of isolation that still lets you get your job done. So, it should be able to read files that it needs um almost full stop. like reading files to me is actually like fine. Now mutating stuff like I I get concerned about and then writing outside of the workspace like I run this in get workspaces so if it changes something
like no problem I can just reset it. Um but that's that's what I look for in sandbox. Um okay so now this is producing a whole bunch of information like were I to go and manually try and find this information myself it would take me a long time but it has already produced this like giant digest of this kind of stuff of what I asked it for. And then at this point, I will just scan this and read this and find out. Um, it going and like giving me this summary already saves me a ton of time beyond what I would have to do.
Um, so and you can ask infinity of these questions. Like it's producing now this like text diagram here. A little bit later I'm going to have it produce uh like an actual like diagram. Um, but like it doing this amount of work just on my behalf and like the amount of parallelization it did to do that is kind of game changer number one. And then in a second I'm going to ask my first um follow-up question about this. So it's going to run its stuff here.
Okay. So this is what I'm wondering. So if it's only exposed locally, how do I off and connect to it, right? So the my colleague like really wants to to run this thing. It turns out that the producers of openclaw um thought about this and so they have documented it and it will the agent here will just tell me about it. So, I'm thinking if I want to deploy this for my coworker, I'm going to go with some balance of a safe option. Um, but also like let them get
their stuff done. So, here they're recommending tail scale, which makes a lot of sense to me. So, this is not on the internet. It's just lock down. You have a tail scale connector and then um you can put that on your phone. So, I would say like I don't know versus like previous me like getting to this point here, I'm already maybe 30 minutes saved. Um, now now I'm asking a follow-up question like security bugs protocol. Um, what's in and out of scope? I'm trying to understand the security model here. Um, and what it's
about to tell me, we have this nice person contact responsible disclosure. There's no bug bounty. Um, but out of scope here, this is interesting. Public internet exposure. I think we've already dealt with that part. Um, prompt injection attacks. Okay, like this is I was not completely new to the idea of open claw last night. Um, prompt injection is like the big spectre here, right? It's kind of a a big spectre on everything in AI is there's no separation of control and data, which by the way, we've made that
mistake so many times in security and we've made it once again like the models are built in a way where they don't understand different trust levels of of who's giving them information. Okay, so now I'm going to say what are we actually like recommending here? Um, presumably the people that created OpenClaw have gone into a lot of thought about like how they recommend to do it and it's going to go and find those details for me and tell me. So we have
this prompt injection reality attack what not to do. Um, producing this kind of level like me reading the docs myself maybe I'm an hour in to time saved or whatever. This this whole thing like this recording here is 6 and a half minutes. So, that's how long agent took to find the details in the first place.
Answer my follow-up questions. All right. So, I'll pause there. I have in total four of these recordings. We'll we'll run through the same way. I want to make sure that we take any questions that folks have like in line as I'm showing these things. >> Yes. Um I guess Naveen asked, "Do you have any skills configured when you're running Cursor?"
>> Uh I don't uh have skills. I a lot of people use them and find them to be very useful. I've seen people show me their skills. I might in the future. um but it hasn't it hasn't been a part of my workflow right now. That being said, I would recommend you know any of these kind of things in AI move so fast and if people are hyping something up, it could be hype, it could also be very useful.
So I'd recommend spend an hour and ask agent to tell you about skills and why they're good and then maybe try one out. >> Okay, >> I have one more question for you. Yes, >> asked like regarding security work like cursor often tries to be helpful and invents fixes and hallucinates which are maybe the opposite of best practices. Maybe it's something that model trained on that's different than your specific codebase um best practices. How would you address this? >> Yeah, 100%. So, anything that I'm going
to produce code for like cursor is going to help me write it but it's going to be me that merges the PR or approves something and so I need to make sure that I understand it for myself. And so if I if I just understand what the model produced, great. Like I review it, makes sense. It saved me a bunch of time. If I don't understand it, then I go back and forth with the model and make it explain what it's doing. And sometimes it's just like, you know what, you're right, I
completely made that up and it'll go and fix it. Um, I will push it down to the level of like showing me docs, showing me code. And usually very very quickly, we get to like some common truth between the model and I. And this is what I I was saying in the beginning like you should just assume that the models are built to like make you happy one way or the other and if it's struggling like it will it will make something up. So have that healthy level of suspicion. That
being said the models are quite good now um and are only getting better. So I what I found myself is the number of times that I need to completely course correct the model has gone drastically down over time. >> And one thing I would recommend trying is like cursor rules. So you can have it automatically and on every model input or also like deterministically based on kind of the context and the description you provide. So for me it was like always hallucinating opening port 3000 even though ours was configured 40001.
So I just have it like always there um as that's something like kind of specific to our codebase. And so think about the security practices if you have kind of a a tenant that you always want agent to know um that's a great chance to check out a cursor roll. >> Yep. Yep. Absolutely. And actually one thing that I've done so I have a I have a personal project that I also use cursor for and what I did is I have a cursor rule that says you know when
you're performing like these basic top level operations go and consult this document that I made for it and then I have that organized into sections and there are explicit details about how it should perform operations like deploy the back end run unit tests and whatever. So that's a good way to like not load your context with very verbose rules but make sure that the agent knows that it should look at them and then knows exactly where to find them.
Okay, prompt number two. So, this is me digging like really into prompt injection here. And again, I just show this for like the level of detail that I'm thinking about. So, I don't need to like I just basically like single shot type this. In fact, it might have typos. Like I don't worry about it so much. So, this is just me saying, "What do you think about this guidance? Copy and paste literally the thing from the docs that it had in the previous window." Um,
and then I'm saying tell me what you think could go wrong. How risky is this? and how uh if at all can we mitigate against it. Um this one is much faster. So this whole session is is about two minutes. But it's going to do its thing. It's going to go and search information, find things. Um I have found that this model opus 4.5 is like very very good at thinking about security stuff. um if you give it the guidance that it should go and do that and further if you if you
tell it um I'll actually show this in a little bit but if you tell it hey when you find security issues I want you to go and do a validation loop like generate a proof of concept and explain the entire chain like end to end um it does an even better job so I actually use um cursor and models as a thought partner in my security work like very often again I need to understand like what it's saying I'm not just going to say model thinks it's a vulnerability.
It's a vulnerability like go do something like but it's going to save me a lot of time in the research. Okay, so here it's listing a whole bunch of um issues like potential issues that I found vulnerabilities and what I'm going to do is actually take one of these. Okay, so um this one I took number one here back up a little bit. Oh, this one.
So this I just wanted to show this is a model hallucination. So it said incomplete sentence the guidance cuts off at prefer mod. I was like you sure like that doesn't make sense to me. I don't think it did that. So I just took it. I said check GitHub copy and pasted the thing and then told it and then now it goes checks the actual source. You're right. I was wrong. So like this is yes model hallucinates. like I just assume it it can and will and I make sure that I understand everything it's giving me and like we we
get to like ground truth here very quickly. So even though the model is wrong sometimes if you have healthy skepticism it's still going to be like much much much faster than like anything that you're doing yourself. All right. Uh we'll take another question and then I'm going to roll into number three.
Yes, there's lots of questions. I have to find a good one. Um, do you find good success using agents to perform security reviews versus just using traditional code scanning tools? Like what's the trade-off between the two? >> 100,000%. Yeah, like agent is so much better and faster. Like I think there is a viable hybrid approach where you take, you know, like a llinter or something and you feed all of those into agent and you have it triage. But the problem with
any kind of like static rules, uh, is that they're static and it doesn't really have context. So it would be very hard to write a good static rule that actually like traced through all the context in the way an agent can. Now it makes things up. So um you have two problems here. You have false negatives.
Um a good way to mitigate against that would be if you really care you do parallel runs and prompting does matter a bit here. Like when you tell agent I'm looking for cross-ite scripting um it's going to be primed to look at that. If you just say like find vulnerabilities, it might not know what kind of vulnerabilities you care about. So like that context matters and I would definitely try different approaches to see what's good and then on the false
positive side putting it through a validation loop. So what I'll often do is I'll I'll have one agent I'll say go find go kick off like sub aents for each of these different groups of of security issues and then four like make a table of the findings and then for each one you have to go and show the full attack chain with code snippets and you have to write like a simulated proof of concept and it will just go kick off it sub agents and then realize like a bunch of
its findings were wrong. But like that full loop would just save me so much time versus finding out things are either unimportant um or false positives in like a normal security scanning model. [snorts] >> And one more question I guess does agent have all your admin credentials and can it go off the hook start using them to do private things? Can they access your environment variables? Um I guess how do you prevent agent from going rogue?
>> Great question. So I am uh a GitHub admin. Um, I'm also uh AWS admin and I actually like one of the things I'm not going to show it here, but like one of the things that I do that's very powerful with cursor is I will actually go and perform AWS actions um with very very elevated permissions.
Now, my trade-off for that is I approve every single one of those and it will write this very complex like AWS CLI, you know, pipe to jq like parsing and I read every single one of them and understand it. Like it's my responsibility if agent does something wrong, it's going to be Travis did something wrong. But it's still so much faster that like me taking the time to review and approve it is a huge time saving. I do not give agent my credentials. So I have here a helper script that I wrote that I tell agent to
call. So we have a prod account. Um if I want to go and get admin, I will say run literally like creds prod and then it goes and calls the script and that script sets environment variables of its shell that agent itself never sees. So now it can run AWS commands that I will approve every single one but it actually like never had the credentials exposed. That's how I do it. Um I think there's like very similar approaches for whatever privileged systems you might
want to be dealing with. Okay, let's do number three. All right. So, is this Oh, this is the other one. Hang on. All right. Number three. All right. Find vulnerabilities. So, this is, you know, to the previous question, how I might look at security stuff. This one's actually very vague. I gave it I'm looking for libraries that might be out of date apps type of issues. If you wanted to be more thorough here, like I have a caveat that don't take whatever whatever we discussed here about open clause
seriously. This is just like a fun thing. If I were doing this for real, I would spend a lot more time than I spent last night running through this. This is just kind of fun demo stuff. But if I were running this for real, I'd probably be more specific. I would go and like read those previous diagrams um in detail and find out the components that I really wanted it to look at and I would probably run like several different kinds of prompts like infrastructure, appsc issues, like specific prompt injection things. But
again, this is just demo purposes. Okay, so this one's about 5 minutes. So, I'm going to pause at interesting points, but we can thread in some more questions while we go. Oh, one thing. So, this has made all of these nice todos, and you're going to see in a second. This is just running through stuff. Like, it's doing kind of like linting level stuff so much faster than like for sure anybody any human could do. If you had a llinter, it could run these equally fast, obviously.
But what's nice about this is the model is driving these llinter type things um intelligently. So, it makes up the right ones to call on the fly and just runs through a bunch of them. All right, another question. >> Yes, I think there's been some interest based on what you previously said. Um, how is it that the agent can see the ENV variables in its shell? Um, maybe talk about the architecture. I think that seems like a initial antiattern.
>> Oh yeah. Okay. Yep. So the um the agent drives a terminal. Um the agent is not like the terminal. It just has a tool. So but it's like it's a property that's set of the tool that it uses. um if if if I ran a thing in the terminal that like exposed it, it's going to see the terminal output, but like the actual like properties of the M are not directly exposed to the agent.
>> And then switching gears a little bit, but when do you use agent mode versus plan and debug? >> Ah, great question. So to be perfectly honest, um initially I was um skeptical of plan and the product team here was very very excited about it. Um, and I I honestly didn't understand because I've been doing these kind of workflows, but actually you'll see in the in the last demo video, I've become a complete plan convert. So, I love plan now. And when I
use plan is I know that we're going to do actually this could have been a plan um if we wanted to, but we're going to build something. And what I really use plan for is for agent to go and like farm for context like more um comprehensively than it would itself. So if we're architecting anything, if I want to build code, any of that, I just by default use plan. If it's like a very quick change, then I won't use it. But if we're going to build anything net new
or make changes to anything that's relatively involved, then I use plan because it will go and like find ahead of time like all of the injection points. I can still ask my questions in plan, but like getting that documented like here's the thing. One makes it form context better and two like keeps it on track like when it's actually doing the execution. And then ask u is that the other one someone asked about ask. >> Yeah. Yeah. So ask is great if like a
lot of times like I just because cursor is fast and the agent's powerful like I will just fire it up and and wonder something completely separate. Um so ask to me is just like I don't want you to do anything except for like go and answer my question. It's a nice um prompt because again when I jump in I'm usually being brief and my ask might have been ambiguous to the agent and it might just start trying to change code and then I I'll tell it like stop you
might not have to do that. So I think ask is like very very useful in those kind of situations. This is just how I use it. Like probably like other people here use ask in a much more advanced mode than I do but like plan um and this basic agent mode are the ones that I use the most. Al by the way um plug for debug. So that was another one that was released. I hadn't used it much, but it really saved me a ton of time like a week and a half ago. So I just think
this this stuff is so cool. Like my recommendation is just go and play with them. Okay, so this one completed and it found a bunch of cool stuff. Um, and I just wanted to show like I'm skeptical of the agent, right? So it's like, oh, we have a P 0. That sounds intense. Like I'm I'm kind of skeptical that like I'm going to jump in and like in 5 minutes of doing AI stuff going to find some like unknown open claw issue. So I'm just I'm lazy. I'm literally like
copy and paste this one and say okay I don't essentially a nice way of saying I don't believe you. Explain it in depth including the attack chain relevant code snippets um and tell me in detail how this would be exploited and prove prove that the exploit works. So this is going to go for another loop now. It's going to do a whole bunch of stuff here and then at the end of the day it's going to say it's not a vulnerability. So it's an intentional feature. The user is
themselves um calling the command here and this would be how I would work my way through um any any kind of validation. We also do security bug finding um internally this way. So we have security bugbot. There are really cool features that are coming that I'm not allowed to talk about today. I really wish I were that would make this kind of a loop on top of PRs and other things like easy for our customers to to customize. So excited for the day when I
can talk about that. Unfortunately can't today but that kind of prompt you know go find vulnerabilities validate them in a loop is extremely powerful and is already finding things that I will not find myself. Like I'm sure if I sat down and looked at the code for four hours like the PR that I could find them, but like realistically like I don't have time to do that level of of of depth for everything. And I think like most security teams are the same. Like I've
never met a security team where every single PR at the company is reviewed in security depth. And now we have these agents that never get tired and are quite capable that we can run on everything. So what it's really done in my opinion is lower the bar on the level of depth that we can cover for all changes. things like threat modeling for any change that were previously untenable are now tenable and I think that's the most powerful part here. Um
other question questions. >> Yes. Can you talk about the different models that you use? Do you typically stick to default mode or do you have specific models that you prefer for different tasks? >> So this is all like a very personal thing. Um I think there are people at the company that that switch models a lot. Uh, I just tend to be a daily driver kind of person. Like I'll get used to a model. I do when new models come out evaluate them. Um, but for me, my daily driver has been Opus 4.5 for at
least the last month, probably longer. Uh, I get used to like kind of I don't know, it's hard to describe, but like I get an intuition for like how the model approaches things and like where it might get stuck in both the pros and cons of it. So there is like some context shift when I try a new one. Um, for example, GPT 5.2 too I found to be very good. Um some GPT models are very very good at solving hard problems but they end up being slower. So like I'll
go and reach for that when I want um in-depth stuff. But this is not you know this is not like the the cursor opinion on anything. This is just like my personal observation. I would encourage everyone to play with any models and come to their own conclusion and figure out where an individual model fits into their workflow. Want to do one more?
Yes, I found a good one. Now I need to go find it in chat. I'll get back to it. You want to >> Okay, cool. All right. So, this is the last demo and then the rest of this is going to be Q&A anyway. So, you you have plenty of time to prepare the uh the good question that you want to ask.
Okay. So, now I'm going to use plan mode and I'm going to say, okay, you know what? This sounds useful. I want to ship Terraform infrastructure. We need to make it safe. Um the developer actually is my boss. They're going to be super mad if I don't get this open claw thing going for them. Uh but also if my boss gets compromised, they're going to be upset. Um and I don't want that. So we're going to make an info directory in the repo. We're going to create
Terraform and we're going to deploy openclaw. We need good defaults based on the docs. Don't make it open to the internet. Um include some plugins that are reasonable. Um, so that you know we can check the box like yeah Travis did the thing but also they're not just going to get like trivially prompt injected. And this is going to be a plan mode. I any kind of like building like this like I said before I would reach for plan mode because ahead of it
starting to do stuff I want to know exactly like what it's thinking and it's going to go like aggressively farm for context up front and by the time it gets around to like this is what we build it will have done a lot more investigation than it would by default. So it's going to go do this for a while.
We can take another question. Yes, I think how would you think about using cursor to make it more secure for all the other devs in the org? So, right now this is like how you do your job, but how do you maybe make cursor secure for the rest of your team? >> Yeah, it's a great question. So I have um I don't know maybe like a controversial or um extremist position on security which is just just my position but a lot of security teams want everybody to understand security is everybody's job. Um I don't feel that
way at all. I think like a great security team like the the biggest gift that we can give our org is actually like handling security problems as much as possible so that everybody else can focus on what they do. So if we're if we're doing our stuff, what I like is most security problems are handled for you by the tools and processes that you use. Those things do not introduce new friction for you. And then you um whatever your job is, like you have a
good intuition and relationship with us where you will come and like ask when something's going on um that you want advice on. And I think we've really nailed that process. So things that I'm very interested in here are putting agent through the right loops so that it can do like self- validation and selfcorrection. Um and then like very early feedback. So like the security bug bot findings that I was talking about. Um we want to as a security team become
involved in changes at the right time where we can help people understand impact. Um and you know like have that careful watching eye on every single change like way more than we are scaled to manually do ourselves. And so that's the kind of like tools and processes we've been focusing on. In addition to that, just um making the systems that are critical to the jobs that everybody do just locked down and secure so that as they engage with them, they can just
worry about what they're doing and not have to think about data exposure and stuff like that. Those are like the services that we provide. Something that we do that's unlike many security teams is we'll handle stuff like patching. Um so our team writes software ourselves and understand software and can take some new vulnerability. we have the right systems in place to make sure that we understand if a vulnerability is actually impactful to us and then if it
is then we will often just go and make the patch to fix it. >> Um and then another question about cursor for penetration testing. Is this something that you do? >> Uh I have not gone very deep in this. I think it's very very interesting. the if you think about you know what I said with uh it finds bugs and then it writes proof of concepts like there is not a giant leap from that to like okay now you have a proof of concept like why don't you run that thing like I think that for sure that's going to happen
like right now the thing that I'm most interested in is using like perfect visibility of the code and the configuration as the source of truth like if I can find a validation like a validated issue then I don't need to have like an actual like exploit done I think there's definitely you chain A to B to C stuff and I know there are teams working on that. It's not a focus of mine right now. I never like to warranty work that I haven't done myself, but I do think it's a very interesting field and people should be
paying attention to it. Okay. Um, so quickly on this. So, uh, I know they're going to want to access email. What do you think I should do from a risk perspective? I'm just going to like cue this up last night as I was doing this. Like I thought of that and so it's it's going to ask me some stuff.
Uh, I'm going to tell it the answers. It's going to keep cooking and then I'm going to cue up my question for it and it will just pick that up like in line. So, I have this question cued up here. It's generating a plan. Hey, here's another thing. Like, agent wrote um this nice diagram, but like it's completely unreasonable to me because it picked bad colors. So, I'm going to go and give it some feedback here. Um, hey, I can't read the diagram with the contrast you
have. Would you mind fixing that? And then it'll skip ahead to here. Now it it has fixed it. It hasn't fixed it yet. It there's a point in this video where it's fixed and I will find but that kind of stuff. It's just like okay model made a mistake. It's very fast for me to go and tell it um it's an issue and now I'm reading the plan of what it's doing. Um this is something that I pay a lot of attention to. Like I'm really curious about like what it decided to do. I will
give it feedback here. Um, so like, okay, what if what if somebody sends my boss like something that uh tells it to ignore previous instructions and like fire Travis or whatever. Like that's the kind of thing that I might ask it. It's going to go do like more validation and loops around this to come up with a good answer. It's giving me details here. I notice here it's giving me like sonnet 4. This is like a training cutoff thing. So like that's what it thought it had at
the time that it was trained. Um, I'm going to tell it to use opus 4.5 instead. it will go and update its plan. Um, and now I've clicked build. So, it's going to go cooking for a while longer. It's going to do stuff. I if I'm deploying this like this is like actually my code and I'm going to understand like every single bit of this. I will use agent to understand bits. Um, but while it goes and builds everything before I get to that, let's take another question. >> Yes, seeing a few folks ask about code
review. Um, how do we use code review? Do we use bugbot? um how do we do you kind of integrate I guess agents in CI/CD? >> Yeah. Yeah, great question. So yes, please use Bugbot. Um it's a wonderful tool. It has saved me from so many issues. Uh it has found cool security things like in in code that I also produced. I use it as like its own integrated validation loop. Um, you know, Bugbot also sometimes will find things that are wrong, but the signal to noise ratio is so high that when Bugbot
has feedback, like it's always worth me paying attention to. And I will just literally like copy paste like I'm I do it the old way. There's actually like a just send to agent and have it fix now, but I will just copy and paste the thing and ask agent to evaluate all the feedback and like see what it does. So, I've always found the the signal noise to be quite good there. And yeah, I mean I love it like agent validating agent. The net impact of this is still going to
be something that I like fully understand, but it has 100% caught on many cases like things that I would have missed myself. >> And then I know some folks are asking about context window. Um how what is your kind of thought process? I know sometimes if you add rules the longer the context window gets it kind of forgets rules. Um do you start new chats early and often? How would you approach this?
>> Yeah, great question. And so this has changed a lot. I think the context window management was like a very big factor when I joined in June and was using cursor. I found it to be much less so now. So I uh about a couple of weeks ago I ran like an all day long session um in I think it was like 9 hours or something and it was so long that like scrolling up even took forever and I didn't reset it once. I found that the compaction feature like its ability to
summarize what's important um is quite good now. So, I haven't found it to be as much of an issue, but you know, if you ever get to like a state where like agent is just totally off on the wrong track, like that would be a good point to tell it to summarize like whatever it's done and then I'll just copy and paste that and spin up a new agent and like paste in the summary and we can get started again. >> Yeah, awesome question.
>> Yeah, I've been seeing a lot of folks interested about plan mode. Seems like that resonated. Um, do you typically when you start with plan um hand it off to another agent to execute? um do you have some running specs others running with it or like sub agents executing just think about like advanced automation use cases? >> Yeah, it's a great question. So I have found that cursor just in executing like it's pretty good about sequencing and like being optimal. Um while it's doing something like like building this out uh
I would just honestly like go on to something else because I'm just going to wait for it to cook. You can use sub aents to speed things up. Really I like uh sub aents more on like a validation loop. Um, and then anytime where it's like if I'm seeing like go find vans across these classes, I'll use sub agent just to like keep them separate and manage context. Um, but yeah, I would say I'm probably not like an expert in this thing. I'm definitely not an expert
in this thing. You should go experiment for yourself and find what works. Like there's so many things. And what's kind of actually what's really cool about this is everybody can develop their own workflow. Like I'm showing you my typical workflows here, but not warranting that they're the best. like if you come up with something that you find quite effective, please share it back with me because I'm always learning too.
>> And then I'm also seeing a lot of questions about kind of what are the enterprisewide security features you should be rolling out. How can you control MCPs, different tools? We're actually having a separate session on that um in two weeks. I'll just drop the link in chat. So I did want to kind of highlight we hear your requests. We have a whole separate session on it. Um would recommend checking it out. >> Awesome. Thank you. Yeah, thank you. I
asked Emily last night. Um, that's a a great session like definitely worth doing. I just want to keep this one focused on on this stuff, but we will have another session to explain all of the security controls and or recommendations for how to keep things safe in an enterprise.
>> Um, so quickly here, so I've as I'm watching it do stuff, I've already queued up questions. So I copy and pasted uh a snippet and I said, "How does this work? Explain the security properties. Show me doc snippets to prove it." Like that's my kind of like keyword for being nice to the agent but telling it I don't trust it and I want to understand. Um, and then also, uh, what are VPC flow log? I know what VPC flow logs are, but I thought it would be
like a great example to go and have show how I will have agent explain something to me. So, it's it's explaining what flow logs are. Um, and now I'm going to tell it to spin a security sub agent um, and look for vulnerabilities in the implementation. Uh, I want to know about any risks before I spin up a PR. It's going to go do that. And then it has some issues, which again, I would I would go and like put through a loop and validate. I want to know if any of
those are like real issues. I can then put up a PR, look at it in GitHub while Bugbot does reviews, um, come to a point where I'm very comfortable with like every single line of code that's being produced, and then I merge it. Like that's basically my end to end rinse and repeat workflow. Um, but everybody has their own. And there's a lot of powerful stuff here. For anybody that hasn't used this kind of stuff yet, like please do. Hopefully I've shown you today that if
you use it well, it can save you a lot of time and be quite powerful. And for anybody that is using it and has workflows that are quite different than these, like please let me know. I I always love to learn. Okay, so that's about it on the video. Let's move to open Q&A.
>> Um, someone asked, do you still write code manually? >> I do not. Um, I know a lot of people do here. So, like being completely honest, like I have always coded. I am not the fastest coder. I think that for my brain and the kind of stuff that I work on, having an agent is like a perfect partner to me because it's quite fast at coding. I am good at thinking in systems and and figuring out where things might break down and I do think from first
principles about anything I build and understand it and think about failure cases. So between the two of us like this very capable agent that writes code very fast and my analytical brain, we tend to do stuff well. If I was a faster coder, I might write it manually still. Although in general like my perspective is that we are moving to a model or a world where agents can write code much faster than even the best humans and we are just going to use them and ask the
right questions to get them to write most of the code. >> And then I guess as you like I think seeing some questions about sandboxes and auto run versus manual approve. um I guess what are the best practices there and how do you think about sandboxing mode and what uh commands should be approved and yeah >> yep so there is also new stuff coming around sandbox um which I don't want to spill the beans on anybody's announcement but we are actively building on it and making it more
powerful um the goal here is to really like give the power to our users and let them pick what they do and don't want agent to run and have the right kind of controls around that what I do um for myself is anything that requires like any um any like network access is going to be an approval. So I want that to ask me every single time, you know, for the reason that like my user an attribute of it is it has more permissions to systems than I'm comfortable just giving to the
agent. I everyone's got their own threat model. Like I honestly don't care like if it RMRFs my hard drive like it would make me a little bit sad but it's not on my top list of like 50 things I worried about. So the the goal is that everybody can set up sandbox in the way that they want it and like whatever your threat model is it works for that. like if you're doing local crypto development or something, you might have a different um threat model than I do. Uh crypto like
the coin, not the not like cryptography. >> You know, we have two minutes left, so apologies for the 60 open questions. We probably won't get to um I'm seeing a lot of questions about cursor rules. Um just best ways to get it organized. I didn't want to kind of highly touch on the highle overview of it. You can have rules that are always applied. We keep recommend keeping those pretty low since they'll be kind of eating up context on every single request. There's rules that
you can apply to specific file patterns. So maybe for all Python files or for all scripting files um adding rules there. And then you can also specify for them to live in different subdirectories. And so on like the main org like the main group, it will always apply to all the subfolders. You can have them live in folders. Um and yeah, I think if you have more questions about this, we do have a cursor 2011 session. Um that's kind of just diving deeper into these
cursor best practices. So would recommend checking them out. Um and then I think another question about bugbot and code review. I think given that it be can take a long time to run. Do you have any advice maybe either ways to have it kind of prompt and fix issues before they happen or yeah what's the best way to kind of combine code review with writing code?
>> Yeah. So I think that you know bugbot in the way I use it is when I I put up a PR it does a review and it's worth me iterating because that's my final step for merging something. Um I think that hooks are also quite powerful. So if you have hook and cursor rules so with a combination of those two like as agent is actually performing actions you can have it do validation and then bugbot really is like a last step and then I also like as you saw will just manually
like drive it sometimes to go find its own bugs uh as I'm developing. So I would say yep, bug bot can be a little slow. You actually would want it to be um comprehensive there. I don't think it's very slow. Like my observation is it's not by a long shot like the slowest thing that we have that runs.
>> Yeah. And I know we are at time. Any advice for all the security engineers in the audience? >> Yeah, I mean just general advice. Um this is so powerful. Like I have been able to It's funny. I got into like a a LinkedIn argument with somebody that didn't believe me, but I have been my self- assessment uh 5x more effective and that was actually like older models like now I'm probably seven and a half 10x more effective and that means that I can do more to protect the business than I could before. And also like new
systems it's like oh I wouldn't have to spend a week to learn about that thing like I can just ramp up and learn so quickly and go and and move the needle like lock it down. like anybody that is not using um agent this way like please after this like go learn how to use it and then if you are more advanced in you know workflows or anything like please share those back with me so that I can also learn um this has been super fun happy to take uh any um you can find me
on LinkedIn or whatever happy to take any questions that folks have um and thank you for joining today thanks for running it Emily >> of course thanks for joining guys have a good one