AI Agents and Security Implications

Hello and welcome back to a special bonus episode of Random but Mmorable. I'm Allie, and I'm here today with Wade and Matt, and we are going to answer some of the listener questions that we did not get a chance to answer last season We also asked for questions recently on Reddit and our community forum. So a big thank you to everyone who responded. And we're going to try and answer as many as possible today. So let's dive right in. First up, we have a question from Caitlin who says, Hi random but memorable. I've been loving the show lately. I particularly love the episode with Glenn Wilkinson about hack recovery question for the team is I've seen some mention of One passassword environments recently and how it can help dev teams. What is the benefit of using it? So What this feature does is essentially move your. ENV files. the secrets within them and make them one password references and then you store them in one password. you can actually sync them. I think that's the coolest feature is like, When we were doing something in the design team and we were like all prototyping and everything like that We just added a bunch of secrets into a into a vault shared them out and then everybody had the right kind of you know things to build. Yeah. so it's basically environment variables and they are usually located in plain text files local on your machine. And then those files sometimes are unintentionally committed as well to a Git repo or a branch or something like that or copy and pasteed into messaging services like Slack to share them and that type of thing as well. So One passassword ennvironments basically helps developers and really anybody using AMO files organize and manage environments environment variables securely And developers can store their environment variables in O passassword and then use the O passassword SDKs and CLI access those values. You can also use one password environments with Cursor and kind of just you know give AI agents secure access to credentials and eliminate the hard coded secrets from your code All right, hope that helped to answer your question, Caitlin On our next question comes from Mike who says Have you ever added a new feature to one password that looked perfect from a security perspective But it failed because it didn't align with how people actually behave I mean, isn't that version one of everything Isn't that like version one of like ninety percent of the things that we do But like when when we launch any feature, we always talk to customers about it And we always have like customer interviews and all of these types of things But like When you test a feature like this, you're usually using a prototype, you're using a beta version, you're using something And like people are outside of their normal environment. They're like they're on a call with a stranger They're like we do our best to kind of put people at ease in these things and we do our best to kind of at for all of these things, but like it's not a perfect environment. it's not with your data. We do betas and we do things with larger audiences as well to try and answer this question. Really, it's always the one point ones that land much better for me. I think you're right about that. I also think that the folks, and I don't want to misspeak, but I do think the folks that tend to have the nightlier beta version of the One passassword extension, for example installed are our one password champions. So sometimes I think features get launched and put through beta fairly well. But then when it goes to kind of genen pop is when we get usability issues with folks that are maybe just not as comfortable as our champions with the product, right, which then helps us iterate to make it more accessible for everyone Mike also asked, if you could eliminate one common misconception about cybersecurity from the public discourse, what would it be My one misconception is that Lking is kind of mysterious and hard It's not really that hard It's like There's enough stuff about cybersecurity on YouTube where anyone could become an expert fairly quickly and fairly easily. And everything has already a program written for you. to actually launch and do these attacks. This probably isn't the best misconception to tell people But cybersecurity isn't as hard as you think it is. That's probably the best one As long as you keep your stuff updated and you have a password manager Like ninety percent of attacks you'll be good from I was thinking the same as you were talking about how easy it is to ask someone. I was thinking that the flip side of that is I think people feel like it's really overwhelming, but at its simplest form, right? It's like have strong passwords, don't use the same one, have multifactor when you can, right? It is simple when you break it down versus thinking that it's going to be complicated. I don't think the misconception piece applies to mine because I think it's already it's already been unraveled in society. But I think for a long time to change your password every ninety days or every thirty days That was everywhere I had security conversations with people and they were like, hey, see what I do to get around this is I use a football team And then I will just move through the players And I showed them how I've been ped. and and like There we were typing in players' names and I was like, look, all of these are probably connected to your to your email address. L it doesn't take It doesn't take an expert to work out what your next password is going to be. like it's probably a Liverpool player. There are not that many of them I've been watching the World Cup, so it definitely depends which team you pick with how strong your passwords are going be. L you pick some Norwegian team other that just their name is super long. It may work out. U But that's a really good one. I completely forgot about that Right? It was big for a while. Yeah, it forever it was We've also had a question from Lachlan Hunt on Reddit who asked Pass keys have been available for a few years now, but many users are still skeptical or actively hostile to them But what do you think has contributed to that negative first impression? and what can the industry do to turn it around I think I don't know what contributed to the negative first impression. I don't know if it was usability issues, but I do think That is how the industry can help turn it around is making it easier to change credential like a type password into a passkey. And I know plug for one passwer, that is something that we do, which I think is great. but I've also seen it on on websites now, right? Like kind of the one click versus having to go through multiple steps I think the The negative factor is probably just the UX of the password is so obvious It's been hundreds of years in use in some sort of like, you know, Roman castle where you utter the password. like it's a common, common thing And then we decided to make it kind of device bound click with biometrics and this kind of thing And the difficulty is just the societal shift And I think we softened that or at least tried to by trying to make it alongside the password for a while. You know, like there's a decent number of websites where you're like, wouldould you like to add a passkey to your account It doesn't overrule your password just yet, but like maybe in the future if we have enough of them, like let's dip our toe in the water. I think that made it more complex, if I'm honest. Like I don't know what they do half the time It adds a pass key to my account and they asks for it ninety percent of the time I'm using one password and it's filled in the password quicker than I could biometrically do anything So a good point. I don't know whether we've made things easier there. I think that we've probably made things more secure. Some sites are using it as two FA as well, which is a bit weird. these companies that add pass keys most of the time are looking to try and get people in quicker because of e commerce and trying to get people to their like shopping carts quicker So like if they can do that via passwords or they can do it by, hey, you were on this device, or like we can just leave you logged in as long as we feel comfortable Or it's the fact that like someone reset their password five or ten times and they're just like really irritated by the fact that they can't remember it. and they don't shop here anymore because of that The retailers are probably really comfortable using both passwords and passkeys That's kind of my take on it. I think like It's going to take a lot societally for us to move on from that I have no rebuttal. That was too good. I just think like the number of passkeys that I have has stayed the same for like Six months Seven months I try to always opt into passkeys. Whenever I see something, I always go to passkeys. but I think you hit the nail on the head with like they're more used for MFA than initial login That is almost what I've seen every single time U The other thing is like the big American banks, none of them use passkeys yet. They all still use SMS text like or just regular password, which I would much rather have a passkey for MMFA rather than SMS text any day Do you think with AI that will change, like how easy it is to And I'm just thinking like fishing, right? Like with the SMS people are a little bit hesitant about that and the voice to confirm your identity. and I know banks do that. Do you think there's any world where as AI kind of takes off and these become more sophisticated, the paskys take off alongside of it. The way the UK banks do it mostly is they require an app now So like I've seen this more and more where like You prove usually via like a video and all sorts of stuff, like that you own your phone And like you're logging in your biometrically authenticated to to your phone. And then when you log in online, it's just like, hey, go and get your phone. Like we're gonna to press a button on your phone and then we'll let you in And so interesting. That's not even pass key based That's like It's a pain because like I'm trying to leave my phone around the house more and more And then like you go and pay for something, it's like, Ah, you're gonna have to go and get it I'm noticing that more places than I'm noticing passkeys is like they will use the phone as the kind of source of truth which is much better than using the email as a source of truth which I absolutely hate because like Most of the time it takes like a good two minutes to come through to my account and then like I don't know, by that time I'm bored. I'm off doing. I thought you were going to say it's better to use the phone because it's more secure, but no, it's actually just a pain. It's just a frression vain. It's just because I'm lazy Okay, finally, we have a question from Victor who says, there's lots of discussion about AI agents signing into accounts and completing tasks on our behalf What role would one password play in this new world Yeah, this is this has become a big thing. I think Apple released some some stuff where AI goes and does stuff on your behalf I think it's becoming a larger part of our online interaction. I actually went to a conference yesterday from Versel And one of the cool things was they showed the number of agents that they have internally that like go off and do stuff And it was Wild I like I thought we had like agents doing stuff like going and getting all research and summarizing it and that type of thing. L Their customer service is ninety one percent. of tickets answered by by a bot that they train on their own software and they found it like more useful than anything. L buying it in externally, anything like that. No, a bot that they've trained on their their material. Wild. So like, yeah, I think that these are going to become a huge thing. like I guess that One passassword's role in ensuring that these tasks are completed securely regardless of whether it's a person or an AI agent doing it This is the space that's constantly evolving. and I think you'll continue to see us come up with solutions for it and evolve it. As of early twenty twenty six, we're solving this from kind of multiple angles. We offer early access to secure agent Autofill, which allows developers to instruct AI agents to sign in with one password instead of sharing credentials directly with the agent cy and paste things and send them to agents use secure agentic autofilm We also offer cursor hooks to connect your One password account to Csor and AI assisted IDE to make coding easier. and we have SAM, the security comprehension and awareness measure. a new benchmark that will test the security skills of AI models and help you avoid using models that fall for security risks fishing and exploitation attacks like that And I for one am really excited to see how this space takes shape I think the number of innovations in this space of how we give agents, only the access that they need and kind of scope it down in all of these sorts of angles for as short a time as they need it is really needed because more and more these agents are kind of going off doing things And most of the time they're doing it on like an API key, which is just like the keys to the kingdom, right Mhm programmed for like speed and getting it done. any by any means possible, right? Like the path of least resistance and as quick as you can. And I think that's when things start to go awry versus No, no, no. this is the lane that you stay in. And for this amount of time and only doing this. Yeahah, I'm one hundred percent with you, Matt. I think that'll be really great to see the innovations that start to come out of that. Did you ever play with any of the like open claw and that type of thing I found That was kind of like the moment for me when I realized If you tell an AI to go and do a job It probably has a wanderr around and has a little bit of a look around the place before it does anything And like I didn't really appreciate that. I was just like, Hey, can you go and like just fill out this markdown file with this thing? And it was like wandering around the folder and it was only when like it's open cllor and it kind of gives you the log of what it's doing that it really like appeared in front of me that was like, o This is like wandering around having a look at stuff. I don't know whether I like that. Like what's it doing with it afterwards? My experience with that, I in the in last year, I had set up competitive agents At the end of every week, they were to email me a summary in a specific format And I noticed that there were outgoing emails from me to people that had some of this it was. so Matt, like I experienced that in December and I was like, whoa And I mean, clearly, there's no instruction about sending emails to contacts or responding to emails that I had received. So that was That was my first experience of it. And I mean, I think things are only getting more sophisticated now, but it was terrifying and also embarrassing. I'm like o Oh my gosh, I'm emailing the teacher, my competitive rundown of, you know. I think when we used to watch sci fi movies and it was like, okay, AI needs to optimize the planet by like removing the humans. And we were like,, this is a good storyline. And now I look at it and I'm like You know what? that probably is like the fastest way like, you know, make make climate neutral. Well, you know, Yeah, it doesn't make it that that much of a sci fi jump anymore For it to when it solves problems in ways that I didn't think it was going to, that's like the truly scary part. I had an agent that I thought could read a repo. I asked, Hey, go go look at all these files, giveive me back the repo and it realized it didn't have access to read the repo did have access to read the Git history So it read the entire kit history to understand what was in the repo And I was like, that was a good bypass. Like I had no I did not even think about that at all So I'm just like, okay, now you know it's in the repo and you can see the getit history. so you can update what you know it's in the repo But it gets crazy when it does just little things like that that you had no clue Do you guys know about the rule of two in Aentic stuff. So there's like three things that AIs can do. They can process untrusted inputs. access sensitive data or private data or change the state or communicate internally, right? So that's A B and C. Uually you only want AIs to be able to to two. So process untrusted data or access data. You don't want change state and communicate. and vice versa. So There's a lot of stuff around detection engineering that's going around detecting AIs that are violating that rule in order to make it so like private data doesn't get sent out or untrusted inputs so like prompt injections aren't occurring within internal communications with sensitive data, right? So a lot of fun stuff around that. Okay, well, although I think we could continue talking for much longer, I will try to wrap it up for us. Thank you for joining us for the bonus episode today. And please stay tuned because we have another bonus episode that is lined up for y'all as well And if you have any questions for us that we didn't have a chance to answer today, please keep sending them to podcast at onepassword. com or of course, comment on any of our episodes on YouTube, and we will definitely make sure to try and answer these during our new season Thank you all for watching or listening, and we'll catch you on the next one.