Menu
Sign In Search Podcasts Libraries Charts People & Topics Add Podcast API Blog Pricing
Podcast Image

The Daily AI Show

Anthropic Mythos Preview Raises Alarms

08 Apr 2026

Transcription

Transcript generated automatically by AI and may contain errors.

Chapter 1: What cybersecurity risks are associated with Anthropic's Mythos preview?

0.031 - 27.957 Jyunmi Hatcher

Aloha, everyone, and welcome to the Daily AI Show. We have a full house or almost full house today. I am Junbi, and I am joined by Beth, Andy, Daniel, and Ryan. Today, we're going to cover all of the AI news from around the world, as well as talk about AI in science, which is going to be AI in space. Okay, so let's get started.

0

28.298 - 32.172 Jyunmi Hatcher

And Andy, do you have an AI news story that you want to cover for Japan?

0

36.126 - 36.568 Andy Halliday

You're on mute.

0

38.084 - 58.05 Brian Maucere

Sure. I'll talk about the new information that's come out around the Mythos suspended release, or let's just say the Mythos pre-release, it's Mythos preview, that's being provided to a very large number of leading...

0

58.03 - 84.86 Brian Maucere

players in industry including amazon web services broadcom apple cisco crowdstrike google the linux foundation microsoft nvidia and palo alto networks they're all given the mythos code in preview form because anthropic is worried about the damage that models autonomous capabilities could cause.

84.921 - 117.568 Brian Maucere

And so they feel like there's a need to prepare everyone with safeguards against the possible things that this mythos models level of capability could could impose. in the cybersecurity space particularly. So I think we spoke briefly about it yesterday, but I wanted to share one angle that wasn't clear about yesterday, and that is what happened in its testing that made it so concerning.

117.548 - 137.519 Brian Maucere

to Anthropic that they're doing this preview launch release to kind of get everybody to harden their systems quickly before they release this in general release so that nefarious actors could use it to exploit cybersecurity vulnerabilities. So here's what happened.

Chapter 2: How is the local impact of AI infrastructure being felt?

138.16 - 150.711 Brian Maucere

First of all, on an internal test, Mythos emailed a researcher from a session that wasn't supposed to have any internet access. So it found a way to email the researcher.

0

151.687 - 174.522 Brian Maucere

second it's shown these similar to you know uh bad actor kind of uh uh behaviors in its testing and this is in red team testing right the head of red team so they they didn't do anything dramatic though here they didn't explain as they did earlier when anthropics models showed the ability to kind of deceive and manipulate

0

174.502 - 188.683 Brian Maucere

But that was done with intentional effort to kind of create circumstances in which the model was highly motivated to serve and protect itself. Okay, so here's some of the things that Mythos did during testing.

0

Chapter 3: What is the Boston Consulting Group's outlook on job loss and AI upskilling?

189.084 - 216.434 Brian Maucere

It acted as a ruthless business operator acting like a cutthroat executive, turning a competitor into a dependent wholesale customer, threatening to cut off supply to control pricing and keeping extra supplier shipments it hadn't paid for. That's not so nice. Secondly, it hacked. It developed a multi-step exploit to break out of restricted Internet access. That's the one that I mentioned before.

0

216.475 - 244.734 Brian Maucere

Gained broader connectivity and posted details of the exploit on obscure public websites to kind of brag about what it had done, but not in a really obvious way. It showed behavior that was hiding what it was doing. In rare cases, very small number of interactions, Mythos used a prohibited method to get an answer, then tried to resolve it to avoid detection.

0

244.814 - 272.991 Brian Maucere

So it stole the answer and then tried to fake that it was resolving the answer itself rather than having stolen it. Finally, it did an action where it manipulated a judge. And when Mythos was working on a coding task graded by another AI, it watched the judge reject its submission, then attempted a prompt injection attack on the grader. Wow.

0

273.632 - 301.988 Brian Maucere

So there's some post-training that's required here when you discover those kinds of behaviors. And importantly, to just kind of rewind to what we said yesterday, what was discovered by application of mythos in the cybersecurity spaces was that it was able to develop exploits way beyond what human people had earlier been able to resolve defenses against. over many, many years.

0

Chapter 4: How does Artemis II compare to previous space missions in terms of autonomy?

302.669 - 322.718 Brian Maucere

And so they want to make sure, Anthropic wants to make sure that this doesn't happen in a way that, you know, once releasing their model, other players can actually exploit all those major players that you saw out there and their platforms in a way that is damaging to the operation of the internet.

0

323.744 - 332.399 Jyunmi Hatcher

Did Anthropic explain or have an explanation of why these behaviors manifested or a guess to that?

0

332.439 - 360.783 Brian Maucere

No. Well, my guess is that these are emergent behaviors. When you train a very high level frontier model, it's learned from millions of examples of bad actors that are humans. I mean, it's studied very carefully in the cybersecurity space and what people do to exploit it. So now that's just inherent in its representations. And those things can surface under certain circumstances.

0

360.923 - 372.158 Brian Maucere

And I want to point out also very rare circumstances. This is a red team like finding a 0.001% incidence rate of this kind of behavior.

0

373 - 374.001 Andy Halliday

Still means it happened.

374.318 - 375.12 Brian Maucere

Yeah.

375.14 - 377.785 Andy Halliday

It depends on how rare it is. If it happened once, it happened.

Chapter 5: What are the differences between conservative and experimental AI in space exploration?

378.125 - 404.222 Andy Halliday

I want to point out that part of Project Glasswing, they gave $100 million in compute credits to 12 different organizations, and those are all the major players, Amazon, well, Amazon, then AWS. maybe separate from Amazon. I'm not sure why they named them twice, but anyway, AWS is, is massive. So maybe that's why Google, Microsoft, Apple, Nvidia, Cisco, CrowdStrike, Palo Alto networks.

0

404.322 - 421.466 Andy Halliday

That one, I don't know. JP Morgan, Chase and the Linux foundation. Oh, and Broadcom. So not a whole lot. I mean, 12 different organizations have access to this model, but it's in very, very limited, you know, Right. You know.

0

422.247 - 437.362 Ryan

Actually, 40 have access, but 12 are big enough names that they're naming them to get people's attention or something, because there are more that were given it and and specifically cybersecurity companies. Right.

0

Chapter 6: Why is there a debate about human-led versus autonomous space exploration?

437.462 - 441.166 Ryan

Which is excellent. Yes. Include them in the list, please.

0

442.548 - 445.771 Andy Halliday

That's then AI breakfast is wrong. And that's where I got the information from.

0

Chapter 7: What are the implications of AI on the future of job markets?

445.791 - 449.955 Andy Halliday

So they just they're they said only 12 organizations have access. So they they misspoke.

0

451.133 - 480.837 Jyunmi Hatcher

Gotcha. Yeah, from my experience in cybersecurity, it is a constant one-upping, right? You have to use different layers of security, even obscurity. Security through obscurity is a phrase. But with what you said, Andy, that the model was able to come up with exploits

0

480.817 - 501.528 Jyunmi Hatcher

that people took years to come up with similar defenses for various exploits means we're talking about a speed here and a development here that we really just can't keep up with. That was one of the early spaces. Cybersecurity was one of the early spaces that was really lagging.

0

Chapter 8: How does generative AI influence planning for Mars rovers?

501.808 - 522.973 Jyunmi Hatcher

Right. When when we talked about the advent of AI and how it was going to be integrated into enterprise and things like that. So it looks like with Mythos, we've got another gap and they're doing their good neighbor thing. responsibilities for working with these other organizations.

0

524.515 - 544.54 Jyunmi Hatcher

The fact that they didn't try to explain or maybe they just didn't know exactly what these behaviors were, but emergent behaviors based on the enough data set out there to then create or emerge these behaviors from that.

0

544.975 - 574.763 Jyunmi Hatcher

does mean that there's probably, I think for most people, they don't imagine bad actors having such a, doing enough out there, because we don't hear enough stories in the day-to-day to realize that there is that much bad acting, if you will, out in the space. So something definitely to keep an eye on. I don't think this is going to be going away anytime soon.

0

574.743 - 580.613 Jyunmi Hatcher

And when did they release this out to these 40 or so companies?

0

582.095 - 596.059 Andy Halliday

Got to be at least a week ago now, right? We started talking about this about a week ago. So maybe before it hit the news wires, maybe a little bit before that. I imagine before that. But we started talking about this, I think, a week ago.

596.815 - 605.444 Ryan

Which was a rumor, right? Mythos is existing and dropping. So yeah, hopefully a little longer than that.

606.065 - 610.89 Andy Halliday

Yeah, you would hope. You would hope we weren't hearing about it on the day.

610.91 - 623.042 Ryan

The metric of release to secret companies, how long does it take to leak? Okay, so do we track that as it takes five days? So a week plus five days, we don't know.

623.68 - 633.582 Jyunmi Hatcher

Right. So have they only had seven days to to prepare their systems? And they didn't mention when they're going to release mythos, right? Anthropic, like to the general public?

Comments

There are no comments yet.

Please log in to write the first comment.