Self-Replicating AI Agents Are Already Here | Logan Graham
MTS
October 17, 2026
Anthropic Frontier Red Team Lead Logan Graham discusses self-replicating AI agents, the state of cyber defense versus offense, and Anthropic's biosecurity initiatives.
Logan Graham
So, the basic premise here is, currently, the world thinks about big, powerful models on large amounts of compute. And there's some bottleneck there, because everybody's clamoring to get GPUs, you have to put them in a big data center, they're super expensive and big.
But we're also getting way better at packing all this intelligence into very small models. And those very small models could be usefully run on small amounts of compute. And then these small models, once they cross this level of autonomy, they don't have to be maybe put there by a human. Maybe they could go find that compute and set themselves up on it themselves. And I think this research is just a first proof of concept of like, yeah, I think we think that future is probably coming.
SPEAKER_2
All right, we are back. We're live with Logan Graham, who leads the Frontier Red Team at Anthropic, which probes advanced and risky capabilities of AI and cyber and in some other fields, which is very topical right now. They just put out some interesting new research on self-replicating AI agents, which is something that has been long prophesied, but we haven't seen it yet until now. So Logan, we're thrilled to have you on. Welcome to MTS.
Logan Graham
Thanks, long time watcher, first time caller.
SPEAKER_2
So tell the audience about what exactly is the Frontier Red Team? What do you do day to day?
Logan Graham
Yeah. So we started pretty early on an Anthropics journey. The whole idea was like, if we think that there might be risks in the future, we should do some weird early stage, very early looking for what those might be. They might happen in three months, they might happen in three years. But the point is, just figure out what the frontier of these risks could be. So if you go back to like 2022, that started with biological and chemical weapons and cyber capabilities back then, basically the models couldn't do anything then. But we were like, maybe at some point, they will. So we need to figure out what to do about them. And then out of that came everything like pre-deployment tests and the responsible scaling policy.
And then this year, I think we've just like, we've crossed the cyber threshold of just, the models can now just do cyber stuff. And now it's like, all these things are now actual real live issues that we all have to deal with. Right.
SPEAKER_2
So how has your work evolved over the last four years? I imagine now it's much more practical and less theoretical. But yeah, in what ways are things really different now in terms of what you do?
Logan Graham
Yeah, I think there are a couple of things. I mean, the first thing is like, everything that we started with, like bio and cyber, you know, like accelerating AI R&D, they're all just real now. But not only that, we just have like large parts of the company and like every lab does to figure out what to do about them. And so we kind of like professionalize dealing with these things. And there's still so much more to go. But now that we've done that, we've set up parts of the company to handle that, like our safeguards team, et cetera. We've been through so many iterations of responsible scaling policy. What we're doing now is like, now that we've hit that, what do we need to care about in the future?
So you see things like our work on replication, for example, or work on multiple agents that are like coordinating with each other, for example.
And a handful of other things like real national security capabilities, recursive self-improvement. What are national security capabilities? So there are, we released a post maybe a couple of months ago looking at some of the capabilities where the models can do things like different surveillance tasks, look at lots of information about you and find out like information that maybe you don't want them to know, as well as just like weaponization. Like we helped to kick off Drone Bench with our loved ones and in labs, and models can now just fly drones.
SPEAKER_2
Yeah, we've had them on end and it's great.
So I'm curious about the state of cybersecurity. I think like one big open question that I have, that kind of makes no sense to me is like, why haven't we seen much more hacking given the capability and the proliferation of like quite cyber capable open source models? Like GLM 5.3 has been out for two months now, and yet we've seen maybe like a modest rise in the amount of hacking incidents. Like EPOC AI put out some data last week that was like, the percentage of Americans that have been victim of a cyber incident has like not increased since the beginning of the year. So like, why not?
Read the other 17 minutes of this episode here.
$15 gets you this episode and 99 more. Credits never expire. After checkout, one link brings you back here to read it.
Not what you expected? Full refund within 14 days.
Thousands of transcripts fetched by people building searchable podcast archives
Fetch transcripts with code
The same credits work from a script or an AI agent. The demo key returns a sample episode in full, no card needed:
curl -H "x-api-key: pt_demo" \
https://spoken.md/transcripts/1000651996090Markdown with the speakers named, for your notes, your knowledge base, or anything that makes HTTP calls.
From $0.10 per transcript. No subscription. Credits never expire. Prices exclude VAT, added at checkout for EU customers. Not what you expected? Email us within 14 days with 20 or fewer credits used and we refund the pack in full.
Using your own key:
curl -H "x-api-key: YOUR_KEY" \
https://spoken.md/transcripts/1000796782049