People Are Getting Weird About AI
Dark Reading Confidential Episode 22: You can love AI without being blind to cybersecurity being sacrificed for utility.
Dark Reading Confidential Episode 22: You can love AI without being blind to cybersecurity being sacrificed for utility.
November 9, 2026
Dark Reading’s Becky Bracken: Hello everyone, and welcome back to Dark Reading Confidential. It’s a podcast from the editors of Dark Reading, bringing you real-world stories straight from the cyber trenches. I am so excited to introduce our guest, Jaya Balou, who is currently CISO and COO of Aisle. Welcome. Thank you so much for joining us, Jaya.
Jaya Baloo: Thank you so much for inviting me, Becky. Thank you.
DR’s Becky Bracken: We are thrilled you’re here. I want to explain to our audience a little bit about why I am so particularly excited to have you here. Recently, I was in the audience at ISC2 Security Congress in Denver, and Jaya gave what I just thought was a blockbuster talk about sort of merging AI and quantum, and it was fascinating. There’s a specific chunk of her talk that I invited her to talk about today. And it’s something I’m just calling: Why is everyone so weird about AI?
And so, it stuck with me for days afterward. I kept having the conversation with people because, as I was just explaining to Jaya, we hear people talk about this on the edges, but I hadn’t heard a technologist and an expert explain it so plainly, which I think Jaya is also sort of your superpower. But just in the past couple of days, we have Sam Altman who came out saying no one needs their own RAM. They’re calling him “RAM Altman.” I mean, it’s just that people are weird. So, can you explain to me why you dropped that section in, and do you know why it stuck with me for so long? Why is it hitting so close?
Jaya Baloo: I think it’s because for ages, when we looked at AI, we had this notion of the uncanny valley, you know? Like, at what point is it so convincing to me that I think I’m speaking to a human? And I think it’s clear to say we’ve passed this uncanny valley. We’re fully, you know, believing the conversations that we’re having with AI. They could be coming from another human. So, the tendency for us to anthropomorphize the conversations, the interactions that we’re having with AI is growing and growing.
And you see people who are not just having a sort of delusional psychosis sometimes of aggrandizement of their own intellect when talking to AI and convinced because of some of the sick fancy that’s built into AI that they’re super intelligent or their ideas are incredibly unique and creative, or that they’re actually the ones to save humanity. You know, I think that the problems that we see with people who are potentially susceptible to that are just made much worse by the current evolution and this current state of AI that we have today. It’s just so real. It’s so convincing. And we then tend to make it more human.
DR’s Becky Bracken: So, let’s go through some of the examples that you gave in your talk. One of them that made me cackle out loud was when you were playing video of a robot doing kung fu with a caption: Why can’t we teach it to do laundry? Explain that example for us.
Jaya Baloo: Yeah, so really what I’m talking about is that, you know, we already have a whole bunch of threats right now when we’re dealing with AI. And that’s really about AI’s capability to both uncover vulnerabilities as well as having vulnerabilities in and of itself. So, AI as a potential for insider threat. As we already have these problems, both in our legacy tech debt and in the new technology that we’re building with AI, we have more problems.
Now, I have no idea who thought this was a great idea to take these things that are entrenched with these problems, and put it into something that can actually cause physical harm and teach a robot, which by the way, I should be clear, this was the Unitree humanoid robot. It was announced a little bit over a week ago. But let’s say that it was announced at the end of October. And this robot was demonstrated on YouTube to do all this cool kung fu stuff.
And the issue with this that I have is that this can cause actual physical harm. What is it actually bringing humans in terms of value? And honestly, I would be much happier with a humanoid robot that can do the wash, fold the laundry, and sort it out in the kid’s hamper. That would be helpful. It wouldn’t be something where I’d be terrified to have this thing anywhere near us. And I think that this is something that I feel is genuinely problematic. Just because you can doesn’t mean you should when it comes to some of these implementations. And if you cannot get the security right, don’t exacerbate and increase the potential for doing damage and harm to humans.
I feel like we forgot that as we’re building. You know, we need to solve tech debt, we need to make sure that we have secure-by-default systems, and then you can implement it in something that coexists with us in the physical realm. So that was my point about that. But I really want to say the real issue here is that Unitree had not one, but two, remote backdoors in its quadruped robots. There are multiple security issues with its humanoid robots. And I think knowing that you can remotely take over these robots, and again, like they’re capable of physical harm, I think that we really need to rethink and really considerately position these types of devices in light of things like the Cyber Resilience Act.
You know, should we allow hardware and software manufacturers to bring out this kind of technology with the potential risk that it poses? And not just from a cyber perspective, but from a physical one as well.
DR’s Becky Bracken: What were some of your other top examples of what’s another one that you’d like to highlight here of people being kind of weird about AI?
Jaya Baloo: Well, so I think a slide that you may have initially seen is, and I kind of didn’t want to reference it right away in my first answer, was about the fact that we’re getting strained in terms of the religiousness around AI. That we both have this tendency to first say, actually, you know, maybe they’re sentient, which Anthropic has been talking about in the Vatican with different religious leaders trying to talk about the potential consciousness of AI.
And although I love science fiction, I love it, I just feel, are we really there yet? Are we there, or is this a marketing ploy? What is this? And I’m genuinely trying to understand because it polemicizes what should be a very rational debate into two camps. Those who believe, like already now, we’re at the point of talking about consciousness and sentience of these AI systems.
And then the other camp, which actually rebukes AI in the name of Jesus. And I think that was the slide that people also laughed at the conference about. But this is real. And I find it bizarre that we have these phenomena happening and that we don’t fully understand the societal consequences either of introducing these systems, which again, let’s talk about what they are. They’re a bunch of graphics cards in a data center running software.
I am a full Kool-Aid drinker of the notion of complexity and emergence and understanding that even though with a relative set of simple ingredients you have potentially complex outcomes. Drank that Kool-Aid long ago. I believe in this stuff, but I just don’t think we’re there yet.
So, I feel this has a sort of reductionism effect for the debate that we should be having over how do we do responsible AI by pulling it into these two polemic camps.
DR’s Becky Bracken: You also tied it to a lot of end-of-times cult-like thinking, but it does raise the question, which you sort of alluded to a moment ago, how much of this is being driven by marketing of these frontier models, of the [Sam] Altmans of the world, of the Peter Thiels of the world who are out talking. How much of a receptive audience are they finding in the cybersecurity sector, do you think? I mean, is that the primary dynamic at play?
Jaya Baloo: Well, look, I think that we need to separate what is actually happening, which is we do have a bunch of vulnerability discoveries that are happening because of AI. We do have automation for real. It is genuinely happening. We do have automation of these hacking frameworks from Strix, RTec-Ai, Cairn. We are seeing hacking frameworks that are actually in RTec specifically, that are actually being used in real-world hacks. So I’m not trying to discount the facts.
However, what I’m urging to do is the call that, for example, Ciaran Martin, the former head of GCHQ (Government Communications Headquarters), always makes, which is, “Calm the FUD down.“
And the fear, uncertainty, doubt that we’re introducing here is not helpful. And the reason that I find it really unhelpful is because we know to a large extent how we should perform some sort of containment in terms of egress filtering and sandboxing of these systems.
They may not be perfect, but we’ve always worked with this notion of defense in depth: assume that one measure fails and that you have additional countermeasures to prevent an actual collapse of the security of the system. I think that that is something that we are actually forgetting right now. And I think it’s really important to say, OK, what are the things we genuinely need that would be considered helpful innovation for this? And I think that what we’re missing is a lot of the observability here around these AI systems.
So, for example, if we’re trying to understand how these systems have potentially been poisoned or Trojanized inherently, then the ability to actually understand how that training data has been populated, even when you only have open-weight models, that would be helpful.
Another thing is additional understanding around this notion of J-space (Jacobian space). So, agent intent. If you know that you have a human operator that brings their own intent to the mix, you know and understand that an AI system might have capabilities and intent that diverge from that human operator, you need to understand when and where that first point of evidence for collecting that divergence could be. We’re not there yet either.
So, I would like us to focus on those things rather than the FUD, like Ciaran Martin and a bunch of other renowned cybersecurity people have been saying for a while, like we really need to focus on the things that are breaking.
And really need to focus on the things that we need to be building rather than everyone up in arms about this “Vulnpocalypse,” “Breachpocalypse,” etc. Apocalypses that are predicted.
DR’s Becky Bracken: What is the best way to push back against this sort of rampant anthropomorphization of these agents? You know, I mean, even when researchers are using words like they will “cheat,” which assumes that there is a moral compass there to begin with. But at the same time, the goal is for them to be as human as possible? How do we thread that needle?
Jaya Baloo: “They will cheat” is the reward hacking notion that this is a utility that is trying to find the shortest path to success. That means that things that a human would consider part of that moral compass, not to cheat, not to lie, not to evade, etc., that we are projecting this somehow on these AI systems and assuming that the guardrails that we’ve put in place will cover it.
But I don’t think that you can make that assumption clearly, because I don’t think we’re always in the position to inspect all of the guardrails and know when they’ve been appropriately applied. And so, I think that there’s also this issue of finiteness. Can you have a finite set of guardrails that are actually encompassing the kinds of things that you want to do? And I think the answer to that is probably not.
You know, you’re probably still going to have a few things that escape your potential for, like having this perfectly constructed set of guardrails. Those are like mathematical limitations.
What I would love to encourage us to do is genuinely try to understand better the types of technology that we put into our networks, give access to our data, give a single API token and say “YOLO” to our enterprises, because I don’t think we have this under control. And I think that’s the real worry that I have.
If we don’t understand fully and we’re not aware of all of the potential risks, for example, cheating or lying, misrepresentation, etc., if we don’t understand that, then we certainly can’t have all of these elevated privileges and permissions that we’re doing. I think this was the point I was trying to make about when you have capability, you need to kind of always have this subset of this accountability and authorization that you’re giving something. And right now, we forgot that lesson. You know, we said, here’s all the stuff that you’ve got access to. You go. And I really feel like we need to rewind a little bit just this part. And that requires better understanding of what they’re capable of in terms of the risks.
DR’s Becky Bracken: So, what is the next step forward? If you had your team and you were seeing some of this sort of extreme thinking, you know, seeping into your defense planning, what would you say to them? What is the correct reset function here?
Jaya Baloo: I’m really worried about the fact that we are genuinely at risk of losing some of the critical thinking when we’re applying AI systems. And you see that already with our coding practices; we are allowing more and more of that initial thinking. So, planning our projects, actually executing the code, reviewing that code, proving it, etc.
DR’s Becky Bracken: Which is shocking in such an evidence-based field. You know, it is shocking that it’s gone so fast. Am I right?
Jaya Baloo: I feel like we kind of learned this lesson in security in the past too. So it’s kind of like, Back to the Future, as well. Every time we tell people about “you should do this because it’s secure,” if it’s not an easy or fun thing to do, you know people won’t do it, right? The utility surpasses the convenience. All of the brain elements that would constitute your frontal lobe warning you for this risk. Like, we just lose it. We become teenagers again. Our risk level and tolerance are high. We are happy to take these risks if it makes our life a little bit easier. And I think we shouldn’t forget that.
So, AI as an assistant is an incredibly powerful, incredibly useful tool. And especially when we’re trying to drive efficiencies with all this technology and all this cost. And you’re talking about, you know, “RAM Altman.” I think the accessibility of AI and the sovereignty of it is not something we should be taking for granted. And it absolutely is a problem.
So, knowing all of this, you know, the drive to get this convenience, to get this utility, to get this thing available, that is so overcoming any type of hesitation we would have, like, “Is this safe? Should I really use this?” “Will I get dumber if I keep doing this all the time and outsource my thinking?” And I think that’s where we actually still need to keep in mind this notion of: What are we giving up to get this utility?
DR’s Becky Bracken: Maybe that’s why what you were saying hit so hard because I haven’t heard a lot of technologists take that position. You know much better than I do, but they are all gas, no breaks, whatever it takes, let’s just not lose the wave. And this return to critical thinking, it seems new. Among your peers and colleagues, do you hear the same kind of concern about losing critical thinking?
Jaya Baloo: I think this is the problem with the Kool-Aid. I think at Aisle we certainly have that. I think there’s a strong tendency amongst security professionals, but also AI professionals. So, my co-founder, Stanislav Fort, who is amazing, has always kind of thought about AI safety and interpretability. And he has been the person to say, for example, we initially set up Aisle to do vulnerability remediation. And you can imagine that as a CISO, I was like, all right, if we can do this and we can do this well, let’s just remediate all the things. Let’s just give it access. And he has always urged caution, always urged caution to say, “You know what? We’re not there yet”
Unless we can fully confidently trust the AI to make those changes, I always want a human approving the PRs to any code request. I just don’t think that what we have in our co-founder is a universal quality in all AI companies. So, I think that like my other co-founder, Ondřej Vlček, who is CISO for multiple companies, including Avast. We have this tendency because we’ve been burned. So we are, you know, a once-burned, twice-shy kind of notion.
We really think about this from a security-oriented risk-mitigation perspective while actually wanting. I have to be very honest with you, I love AI. But that love shouldn’t be blind. It should take the appropriate measure of what is it that I love? And why do I love it? And, you know, what is the benefit to humanity? And Becky, there’s a story that I didn’t tell on stage.
There are so many good things we can do with AI, but the one thing that I’m the most encouraged of is … have you heard about the story with the guy and the dog in Australia?
DR’s Becky Bracken: No, I don’t think I have.
Jaya Baloo: So, this was a machine learning engineer whose dog developed cancer, inoperable, and the dog had to be put down, was the diagnosis. And what he then did is sequence the dog’s DNA. You know, this is a guy with no biology background, sequenced the dog’s DNA, and figured out what was wrong in the DNA that caused the cancer, developed an mRNA vaccine with AlphaFold.
The most expensive part of the experiment apparently was to actually synthesize the vaccine. And then he administered it to his dog; he had to get permission for that. And Rosie’s tumors shrank, and she’s is still here and is still hanging out and playing ball with this machine learning engineer.
So I love this story, and this is why I love AI. Because I think if it can unlock the potential benefits, that is amazing. I want those benefits. I just want to make sure that we minimize the downside risks that are inevitable with this kind of technology. And we just need to figure out how we do that. So, we shouldn’t go screaming into the abyss, like, yay! You know, I don’t want us to be a bunch of lemmings following AI down the abyss.
I just want us to manage that risk and take the benefits while eyes wide open, understanding that right now we need to make sure that we do have enough RAM, you know, and that we do have sovereign compute, that we do not sacrifice all our clean water
DR’s Becky Bracken: No one has to give up their computer for OpenAI?
Jaya Baloo: Yeah. I think the future of AI is probably more decentralized than the frontier labs have us think right now.
DR’s Becky Bracken: How much of this is being driven by investment dollars and hyping up stock price? Because we talked around in our editorial meeting about well, is this just a collective psychosis issue? And it’s like, well, first, let’s be careful about saying “psychosis,” but also are we being marketed into a psychosis?
Jaya Baloo: I would love to have a balanced approach. I’m not saying that there’s no marketing being done. I think that is certainly being done. I don’t think we should eliminate that. I think it would be naive to assume otherwise. At the same time, going full tilt that it’s all just marketing is the wrong approach. I think two things can be true at the same time. And I want us to consider this more nuanced thing. And I really want our brains to hold the capacity that both things are true.
There is absolutely a marketing effort. OK, when you have these kinds of companies going to IPO and it’s not just OpenAI, it’s also Anthropic, and it’s also probably a whole bunch of other, you know, parts of this wheel that are actually driving this rage. But there are genuinely risks that we must counter. So, I really do worry about that. And I don’t think that that is psychosis.
You must have seen the filing of Anthropic for the IPO. And one of those things that is in that document is there’s a very real existential threat that AI could kill us all. I think that’s maybe the first time that was ever put into a filing of a company potentially going to IPO. And even there, I think, surely that is not marketing. Surely there are intelligent, thoughtful humans that want to warn of a potential risk that they genuinely see for themselves. I think we need to evaluate the difference between that and how much space there is between the marketing and the genuine risk. But I do think there is a genuine risk. There is. And we shouldn’t downplay it as psychosis, while at the same time recognize that there’s marketing.
DR’s Becky Bracken: I want to thank you so much for coming. You’ve given me a whole other set of things to think about. Thank you for sharing your deep expertise and your thoughts with us today. We really appreciate it.
Jaya Baloo: No worries, thank you for having me.
DR’s Becky Bracken: This has been Dark Reading Confidential. It’s a podcast from the editors of Dark Reading, bringing you real-world stories straight from the cyber trenches. I’m Becky Bracken. Thank you so much for joining us. We’ll see you next time.
