00:00Let's start with what this is. Can you briefly describe what is Mythos?
00:04So everyone who has used an AI chatbot such as ChatGPT or Claude is familiar with that kind of interface.
00:12And over the past year, we've seen dramatic progress in these capabilities,
00:16not just in chatting with users, but actually writing software programs,
00:20and as they get more agentic, building fully fleshed out systems.
00:24So if you're really, really good at writing computer code,
00:27it's not that far away to be really, really good at hacking computer code.
00:31And what we've seen is an explosion, a qualitative shift in just how much these models can achieve.
00:38This is the latest and greatest anthropic model,
00:41and it is capable of executing not just finding a bug, but exploiting that bug,
00:47going through all the links of the cyber kill chain to successfully execute attack and doing so autonomously.
00:54Now, why does that matter?
00:55Because even though we're talking about computers,
00:58cybersecurity has actually been under a massive labor shortage for the past 20 years.
01:04If you think about, you know, why aren't we able to patch all of these bugs,
01:09it's because hackers are really expensive.
01:11These are talented folks who are difficult to train.
01:13But now, if AI can do so much of that job by itself,
01:18you start thinking about Anthropic as a company having capabilities comparable to that of the National Security Agency in the
01:26United States government.
01:27And fortunately, Anthropic was the one to get to this future first.
01:31They're approaching the United States government.
01:33They're approaching critical infrastructure companies.
01:35They're approaching the key cloud providers and banks and saying,
01:39this is what we've discovered.
01:40You live in a different world now.
01:42Let's figure out how to secure it.
01:44From your standpoint, are you fearful that this could expose vulnerabilities or hopeful it would improve resilience?
01:48So I think there's a sequencing part of all of this story.
01:52This is a capability that is useful for cyber offense.
01:55And that means doing bad things, but also doing things that are critical to U.S. national security.
02:00But it's also a boon for cyber defense,
02:02because you can imagine every new piece of software before it enters the world is going to have to go
02:08through the mythos test,
02:10which is to say, can you survive this gauntlet?
02:13And the reality is, going through that gauntlet before was, number one, pretty expensive and time-consuming,
02:20and number two, not perfect.
02:23There was a lot of bugs that weren't discovered,
02:25and we know that because Mythos just discovered a giant pile of new bugs and all this stuff.
02:30So this does offer a world in which we can secure our software in a bigger way.
02:37The problem is that transition point.
02:39There is a massive amount of code, whether that's open source code,
02:43and it turns out that much of the infrastructure of the Internet relies on stuff that is maintained by volunteers,
02:49and also big critical infrastructure providers.
02:53We're talking banks, energy companies,
02:55and the amount of overhauling that they have to do over the next 12 months is massive.
03:00And the question is, you know, how long can we keep the Mythos capability exclusively in the hands of the
03:07good guys?
03:07How much can we patch before problems start arising?
03:11How far behind is the government from the private sector,
03:13and how willing is the government to work with said private sector that there's been some tension between?
03:18So if you're talking about could the United States federal government create something like Mythos?
03:24Could they create something like a frontier AI model?
03:27They are very, very, very far behind.
03:30They're really not even trying.
03:32So the only way that they can get access to these capabilities is to work with private industry,
03:39the types of folks who can do this.
03:41And in this case, Anthropic has what they estimate to be a 6 to 18 month lead,
03:46even over open AI, even over all the other best AI labs in the world, which are in the private
03:52sector.
03:53So that period of time is a critical window to start patching all of this computer code
03:59before China catches up, before hacker communities get access to comparable capabilities.
04:04And tragically, that window of time is one that the U.S. government is at risk of squandering
04:09because of a dispute over contracting terms in Anthropic's working with the Pentagon.
04:15Just for folks who aren't aware, Anthropic is a critical provider not just of cyber AI capabilities,
04:21but also of AI capabilities that are useful in intelligence and warfighting,
04:26literally being used in Iran right now every single day.
04:30And despite Anthropic providing all those critical capabilities to key national security agencies,
04:35over this contracting dispute, the Pentagon decided to label them a supply chain risk.
04:40The type of designation that we use when we find out a company is secretly like a front for Chinese
04:45spies.
04:46It's completely unreasonable.
04:48And the government is in this moment of utter confusion.
04:53Because on the one hand, they're saying, Anthropic,
04:56you've got to help us secure the financial arteries of our economy on cyber.
05:01On the other hand, we don't trust you at all.
05:02And actually, we'd prefer if you go out of business.
05:04When we talk to different executives and we ask them about mythos,
05:07they say, well, there's as much good as there is bad.
05:09And we're actually really optimistic.
05:11When you talk to them, not on a camera, what's the level of concern right now
05:16among executives as well as government officials about what this could do in terms of systemic disruption?
05:21So I think the way to think about it is what if Anthropic wasn't going down this path, right?
05:27Anthropic is suffering in business terms because they're not offering mythos to all of their customers right now.
05:34They've made a strategic decision to withhold these capabilities,
05:37to invite in 40 key companies, to invite in banks and government agencies to try and secure the Internet from
05:44this degree of cyber power.
05:46In the alternative scenario, where this fell into the hands of somebody who wanted to use it to make bets
05:53on the market,
05:53or somebody who wanted to use it to exploit cyber offense in a national security sense,
05:59maybe one of our adversaries, we would be having a completely different conversation.
06:05So the executives are probably right.
06:07Because Anthropic has chosen this strategy of engagement and support, the long-term outlook is pretty good.
06:14What they're really going to be dealing with is an incredibly messy middle.
06:18Gregory, you said that Anthropic is being used right now when it comes to the war.
06:22How has Project Maven changed how the U.S. fights war?
06:25So Project Maven started out as a relatively humble project to incorporate AI into analyzing drone surveillance feeds.
06:34So if you're familiar with those predator eyes in the sky type of drones,
06:39it turns out that here, too, there's a major labor shortage.
06:42It turns out that the U.S. military has a really hard time getting people to watch all of those
06:47screens simultaneously.
06:49So Project Maven thought we can bring in AI computer vision.
06:52They can help these analysts by doing sort of the first processing pass of this type of footage.
06:58And that was back in 2018, 2019.
07:02What it could do is it could say, hey, in all of these satellite photos or drone photos,
07:07this part is all empty water, but right here there's a warship.
07:11Fast forward to today, and the capabilities are much more sophisticated,
07:15in part because we're incorporating things like Anthropic models, large language models.
07:21And what that means is that the AI can no longer just say, here is a warship.
07:26It can say, hey, there was no warships in this area yesterday.
07:30Now there's 10.
07:31They appear to be headed in this direction.
07:33And it can write the first draft of the intelligence report about what those things might be up to.
07:38And it can speed up all of these workflows.
07:41When you hear about the fact that, for example, the United States military struck 1,000 targets in 24 hours
07:49in the first part of the Iranian war,
07:52you know, the benchmark in the U.S. government used to be like 50 per day.
07:56And the key thing is not just how many things you can strike, but the fact that many of those
08:01targets are mobile,
08:02many of those targets are disguised.
08:04So when you hear about 1,000 struck targets in Iran, what that actually means is 10,000, 20,000
08:12high-quality intelligence processing
08:14and targeting decisions that had to be made.
08:17There's no way the U.S. government could have done that without the help of AI.
Comments