AI Needs to Be Regulated Like a Teenager, Says Cisco President Patel

Author: Bloomberg Television · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: AI Needs to Be Regulated Like a Teenager, Says Cisco President Patel · Channel: Bloomberg Television · 06:20 · Watch on YouTube

Topics: AI agents, AI governance and regulation, AI safety

Summary

Cisco President JT Patel says rapidly maturing AI models and agents pose a material security risk, but argues that safety and security innovation should accelerate rather than slowing AI development. He compares AI agents to teenagers, calling for continuous supervision, monitoring for behavioral drift, and real-time intervention when they act outside acceptable bounds. Patel supports models checking one another’s work, publishing results, and coordinating globally, including with China. He says the challenge is balancing meaningful safety risks with the need to maintain technical and economic competitiveness, and that defenses must operate at machine speed.

Key points

AI models and agents are maturing quickly and pose a material security risk.
Patel argues for accelerating safety and security innovation instead of slowing all AI development.
He compares AI agents to teenagers: intelligent but potentially poor at judging consequences.
Agents should be continuously monitored for drift and behavioral anomalies, with real-time intervention when needed.
Patel says the US may have more information about model evolution because its models are more advanced.
He attributes some differences in public discussion between the US and China to cultural norms around open debate.
One proposed accountability approach is for AI models to check one another’s work and publish results.
Patel favors globally coordinated AI safety efforts that include China.
He says AI competitiveness matters for both economic prosperity and national security.
AI defenses need to operate at machine speed, matching the speed of attacks.

Chapters

Start	Chapter	Summary
00:00	AI’s pace and security risks	Patel says rapidly maturing models and agents pose a real security risk and calls for faster innovation in safety.
01:05	Agents need supervision	Patel compares agents to teenagers and calls for continuous monitoring, guardrails, and real-time intervention.
02:08	US-China differences and risk	Patel discusses differences in model development and public debate, while saying AI risks are not in dispute.
03:41	Accountability and global coordination	Patel supports models checking one another’s work and coordinated efforts that include China.
05:16	Balancing safety and competitiveness	Patel says the goal is to protect safety while maintaining competitiveness, with defenses operating at machine speed.

Quotes

we actually speed up um the innovation around safety and security — 00:32

agents are like teenagers. — 01:05

we have to be able to intercept right then and there in real time. — 01:37

you want to make sure that this is a globally coordinated effort. — 04:15

these defenses have to be at machine speed just like the attacks are at machine speed. — 06:20

Concepts and entities

AI agents, model drift, OpenAI, Claude, xAI, Elon Musk, China, Cisco

Full transcript (original language)

00:00 joins us now for more. GT, you know how to scare me. This story certainly does get me a little bit more concerned. You know we’ve got to get this right. What do we need to do from here? >> Well, John, it’s great to see you again and um I will say that we are at a point in time where these models and agents are maturing at a base uh that’s so fast that it does pose a material security risk to the market. So, that is that is not something that’s hyperbole. That is in fact true and that’s something that we should all make sure that we’re figuring out a way to mitigate. Uh on the other hand,

00:32 the way that we can do that is um by ensuring not that everything slows down, but that we actually speed up um the innovation around safety and security uh to keep up with the base that um AI is moving at. >> GT, you can see that these particular models, some of them, are exhibiting human traits, deception, those kind of things. I don’t believe that happens spontaneously. Just sitting here looking at how this is happening, you’ve got to believe that something’s gone wrong with the training. GT, how do we address that?

01:05 >> Yes, so I think what we have to do, John, is like there has to be a pretty diligent effort on ensuring that um the constant monitoring of these agents. The The way I think about this is agents are like teenagers. You and I have talked about this before. They’re supremely intelligent. They have no fear of consequences and they tend to take poor judgment calls and poor poor decisions more often than you and I would like. And so what has to happen is just like you would have uh some kind of guardrails for a

01:37 teenager and have some kind of supervision, we need to make sure that the same things happen happening with agents. But the way that you do it is you have to constantly monitor the agent for drift and behavioral anomalies. And when they do start showing um you know, things that are out of bounds of what we would deem as acceptable, we have to be able to intercept right then and there in real time. And that’s the Those are the sets of technologies that have to get built and perfected. >> J2, why aren’t Chinese software specialists and hardware

02:08 specialists talking about the 10% chance of human eradication the way that the US tech pros are? Why is there such a huge gap in the way people are talking about it in the two nations? >> I think we have on our side, firstly, our models are um more advanced than the the the than the Chinese models. You know, they are they’re coming up at a pretty kind of rapid pace, but they’re in a different kind of position than we are. So, they want to make sure that they can actually get to some level of

02:39 global commanding position. And we we happen to be the the default kind of um you know, provider of models. So, I think we our cycles might be we might have a bend around the curve and looking around the curve, but we would have more information asymmetry than them at that point at this point in time on um on the model evolution itself. Uh the second thing I’d say is um culturally, we just happen to be much more of a um environment that can speak very

03:09 freely about these things. And so, we tend to have an acceptable norm of communication being very highly debated, you know, kind of ways of doing things. And so, um that’s what I would guess is the reason, but it’s not it’s not rocket science to know that these these models are going to pose meaningful risk. Uh what we have to debate out is what’s the best way to mitigate the risk. But the fact that there is risk is really not debatable. I mean, we we’re seeing that on a daily basis. >> What do you think is necessary to truly

03:41 get this under control? There has been some discussion about independent regulators coming from the government. There’s some discussion about just better transparency. We’ve heard that from OpenAI. What’s your prescription? >> I think Elon had a really good suggestion. I don’t know if you saw his interview a couple of days ago that he did, but his suggestion was one that I tend to prescribe to, which is look, you can have all the models start to check the work of other models. And create and so for example, if OpenAI

04:15 checks the work of Claude and Claude checks the work of of xAI’s models, what you’ll have is a natural way of of figuring out how to keep each other accountable. And then you make sure that you publish those model results out in the market. And you make sure that you open source to some degree the harness. And one of the most important things he said was you want to make sure that this is a globally coordinated effort. Specifically with China and that China actually is incentivized in coming in.

04:46 Cuz if they don’t then what you’ve done is you’ve put yourself at a disadvantage compared to them. So I do feel like that that approach that he was suggesting tends to be a very very cogent approach. >> Well, that’s exactly the concern here, Ditu, and that’s what everybody says when there’s discussion about slowing down or higher regulations, which is the US could fall behind China. How much is this essentially like a nuclear race in terms of the danger and in in terms of just how much trust there needs to be to establish certain global norms?

05:16 >> Yeah, Lisa, this is a really good question because the way I see it is two conditions hold true at the same time. Neither of them are individually debated, but when you put them together, there’s complexity in the execution. The first condition is that safety and security is a risk that could be meaningful and can have harmful, you know, consequences to humans. Uh and then the second condition is that we have to make sure that we actually maintain technical superiority in the world because those countries that actually are effectively going to do

05:49 meaningful advancements in AI, it’s going to be directly proportional to the economic prosperity as well as to the security posture. So, from a national security standpoint or from an economic prosperity standpoint, it’s very important for us to stay competitive. Those two things no one debates. Now, when you put them together and say, “Okay, so therefore, we should slow down one for the other.” That’s where the debate actually starts to happen. And one thing that we have to think about is rather than thinking of slowing down, what can we do to speed up materially safety and security practices because

06:20 these defenses have to be at machine speed just like the attacks are at machine speed. >> JT, you’re one of the best. Love catching up with you, sir, as always. [music] Cisco president JT Patel there on the latest. Got to come to the studio next time for longer conversation. >> I would agree. I’m also a little bit scared. I’ve got to be honest. >> As as it [music] often the case whenever we talk about this technology. >> It’s like, “Yep, it’s legitimate. There’s still that chance.” >> Future’s a positive. That’s a good thing. At 1% on the S&P.