At this point, the secret is out that the quickest path to a laughably large stock boost is simply uttering the term “A.I.” On Tuesday, Allbirds, the formerly almost-cool shoe company, announced that it’s making a slight pivot: It is now, no joke, getting into the business of A.I. compute and G.P.U.s-as-a-service. The new company, which had lost 99 percent of its value since its I.P.O., will call itself “NewBird A.I.” And no, it’s not like they happen to have a glut of G.P.U.s on hand—they’ll have to buy the chips first. So, another good day for Jensen Huang.

Shares in Allbirds surged as much as 800 percent on Tuesday, reaching levels the company hasn’t seen since 2024—and they’re still up more than 370 percent since last week. If I were going to call a bubble peak, this might be it.

And now for the main event…

Seven years ago, back when OpenAI was still a nonprofit and Sam Altman wasn’t yet a household name, the lab trained a 1.5 billion–parameter model it christened GPT-2. At the time, the team was struck with terror by their own creation. Concerned about its potential to generate “deceptive, biased, or abusive language at scale”—which, in retrospect, seems pretty tame—OpenAI opted not to release it. Instead the company provided a less powerful model to researchers, shared a bunch of technical details, and nodded gravely to the industry’s growing ethical responsibilities. Dario Amodei, who at the time was still heading up research at OpenAI, was part of the team that made that decision, and explained the motivation to the tech community.

So it’s kinda hard to know how seriously to take Amodei now that he’s essentially rerunning that release strategy for Mythos, Anthropic’s new model, which is rumored to possess a few trillion parameters and, the company claims, is simply too dangerous to be released. Plenty of people in the industry are justifiably skeptical of the various horror stories about Mythos that have been leaking out in the media—like the almost-too-perfect anecdote about how it escaped from containment during testing, gained its own internet access, and emailed an Anthropic researcher who’d been eating a sandwich in the park.

But the risks Mythos presents, at least according to Anthropic, are very real—a true “clear and present danger” in cybersecurity, per a post Amodei wrote on X. And, at least on the surface, the company appears to be treating the situation with deadly seriousness. Rather than release the model to the public, Anthropic launched an initiative called Project Glasswing, providing a preview version of Mythos to a group of 40-odd organizations, including the likes of Amazon, Apple, Broadcom, the Linux Foundation, CrowdStrike, and so on. Other than Cisco—which acknowledged that it is “integrating Mythos into its security and development workflows”—none of the listed firms returned requests for comment about their findings. (Though sources told me that researchers at a few of those organizations were not too impressed.) As part of this initiative, Anthropic has also engaged the U.S. government, despite the Pentagon’s recent decision to ice the company out: Bloomberg reported earlier today that the White House is planning to make a version of Mythos available to some federal agencies.

The goal, according to Anthropic, is to give the world’s top tech companies and governments time to fortify their cybersecurity defenses against the “thousands” of “high-severity vulnerabilities” that Mythos was able to identify and potentially exploit in critical software around the globe. Some of these included decades-old bugs, often ignored or overlooked, in a couple of open-source software projects. Anthropic is also giving Glasswing testers $100 million worth of Claude credits so that participants—including some competitors, like Google—can use the system to identify vulnerabilities in their own code. A philanthropic gesture and a P.R. coup, all rolled into one. Anthropic did not respond to a request to discuss the details of the Mythos announcement.

There are plenty of reasons why various critics have panned the release as marketing hype in an industry already teeming with it. Notably, Mythos’s system card spends a mere five pages on the topic of cybersecurity, while devoting nearly 40 pages to assessing the model’s psychological “welfare” and another dozen pages to pondering whether it’s conscious. Indeed, Dr. Heidy Khlaaf, the chief A.I. scientist at the A.I. Now Institute, noted that Anthropic chose not to disclose how often Mythos “found” vulnerabilities that weren’t actually there (false positives), and that it didn’t describe how human evaluators rated the system’s bug-finding abilities. Independent researchers can’t do this work or compare results with Anthropic’s either, since the model hasn’t been released.

That doesn’t mean, though, that the model’s cyber capabilities aren’t real. The A.I. Security Institute in the U.K., the only organization so far to conduct an independent and public evaluation of Mythos, found it to be “a step up over previous frontier models.” In controlled evaluations, Mythos could autonomously execute multistage attacks and exploit vulnerabilities—work that would take days for a human professional. Mythos was also the first model AISI tested on a 32-step simulation of a cyberattack—called a cyber range—that managed to complete the simulation, although it succeeded only around a third of the time.

And even that pass rate is rife with caveats, according to Boyan Milanov, a cybersecurity researcher at A.I. Now. The range, Milanov told me, was “small, weakly defended, and vulnerable”—which means solving it doesn’t necessarily translate to real-world attacks—and even this weak range stymied Mythos seven times out of 10. But others have more faith in the model’s capacity for carnage. Kanishka Narayan, the U.K.’s minister for A.I. and Online Safety, said on X that “Mythos is the most capable model we’ve ever evaluated for cyber. This represents a step-up in A.I. cyber capabilities.” His urgent message to businesses: “Get the basics right now.”

When I called up David Kennedy, a former N.S.A. hacker and the C.E.O. of the cybersecurity firm TrustedSec, he acknowledged that “a lot of the Mythos stuff was very, very overhyped.” But he did underscore the ability of increasingly advanced L.L.M.s to empower bad actors. To wit: He posited that existing models, accessible to the public today, could be used to conduct intelligence-gathering and reconnaissance on a company; trawl through LinkedIn to find specific targets; craft convincing-looking webpages designed to harvest sign-on credentials; fire off phishing attempts to get people onto those sites; highlight avenues of access to a company’s system; etcetera. If you’re a cybercriminal, A.I. models are “gonna up your game 1,000 percent,” Kennedy said. “You really don’t even have to understand how to hack to actually hack. So that’s obviously super concerning.”

This trend has been playing out since 2023. According to SentinelOne, global cybersecurity data breaches have increased by 40 percent so far in 2026—that’s about 2,000 attacks per week. In 2025, according to IBM, 16 percent of reported data breaches involved attackers using A.I. tools. CrowdStrike’s recent Global Threat report noted an 89 percent year-over-year increase in attacks by “A.I.-enabled adversaries” in 2025. These statistics aren’t abstract, either; earlier this year, a single hacker leveraged A.I. tools to successfully breach nine Mexican government organizations, accessing hundreds of millions of civilian records. Research by the cybersecurity firm Aisle found that cheap(er), open-weight models could identify almost all of the same vulnerabilities that Mythos did.

At the same time, however, the limited information we have about Mythos suggests it could supercharge this trend. Aisle C.E.O. Ondřej Vlček told me that the most impressive (or concerning) thing about Mythos is not its ability to identify vulnerabilities, but its apparent ability to autonomously exploit them. Like the rest of us, he’d like to be able to independently verify Anthropic’s claims. But he said that if the description of the model’s capabilities is true, that’s a reason not to release it to the public. “That is a functionality that is only useful for the bad guys,” he said.

Alas, the game theory of the A.I. arms race dictates that if you’re not competing, you’re getting left behind—and there will always be some market logic to justify staying ahead of your rivals. On Tuesday, for instance, OpenAI released a version of GPT-5.4 that essentially pares back the model’s guardrails to enable more-potent cybersecurity applications. Since this makes it more dangerous, OpenAI, like Anthropic, has launched the version to only a limited number of organizations and researchers, for starters. (Though this release doesn’t have a fancy name like Project Glasswing.)

Meanwhile, a source told me they doubt Anthropic is the only company to have a model as cyber-capable as Mythos—two to four organizations likely have something similar, this person speculated, including Chinese companies, and someone will likely open-source one of them soon. Vlček noted that this moment should be a “big wake-up call” to organizations to take cybersecurity more seriously.

Kennedy, for one, is absolutely sure that Mythos will be publicly released, eventually. But he’s less concerned about Anthropic’s release schedule than he is about the risks he sees coming from inside the proverbial house. Efforts to adopt and integrate generative A.I. systems at scale across enterprises are continuing, and each integration opens up new threats and pathways for exploitation. “The problem with A.I. is that it doesn’t have any security boundaries at all,” Kennedy said. Agentic integrations—which strip back pretty much all the security around online permissions and sign-ons—open up external access to entire code and file bases. “No security program I have seen today, in all the companies we work with, has anything in place to be able to handle that.”

Due to the steady proliferation of A.I. models, Kennedy expects the volume of cyberattacks to increase “astronomically” this year. “The entire hacking community, on the illegal side, is 100 percent all in on A.I.,” he says. “People aren’t looking at the techniques that hackers are using right now. They’re not looking at where A.I. is really heading. And I think we’re drastically unprepared to handle that.”

That’s all for today. I’ll see you next week.
Ian

Need help? Review our FAQ page or contact us for assistance. For brand partnerships, email ads@puck.news.

You received this email because you signed up to receive emails from Puck, or as part of your Puck account associated with {{customer.email}}. To stop receiving this newsletter and/or manage all your email preferences, click here.

Puck is published by Heat Media LLC. 107 Greenwich St., New York, NY 10006