Is Anthropics Claude Mythos a big stunt, or a real security threat? What the experts say.

Anthropic put the entire tech world on notice last week with an unprecedented announcement: it made an AI model so advanced that it was too dangerous to release to the public. Anthropic said the new frontier language model, Claude Mythos Preview, would "reshape cybersecurity."

Anthropic also announced the formation of Project Glasswing, an invite-only group of organizations — including some of Anthropic's biggest competitors — to test Claude Mythos Preview and secure their infrastructure.

Anthropic said that Claude Mythos Preview "found thousands of high-severity vulnerabilities, including some in every major operating system and web browser." (Emphasis in original.) The company said Project Glasswing was necessary "to help secure the world’s most critical software."

By Friday, CNBC reported that Federal Reserve Chairman Jerome Powell and Treasury Secretary Scott Bessent had summoned the high priests of finance (aka banking CEOs) for an emergency meeting about the new model. New York Times writer Thomas Friedman fretted over a "terrifying" future in which any teenager armed with Claude could hack the local power grid.

The reaction to Claude Mythos Preview quickly split along predictable lines. AI boosters hailed the new model as proof that artificial general intelligence (AGI) was nigh, praising Anthropic for rolling it out so responsibly.

Critics and AI skeptics called Project Glasswing a big publicity stunt.

So, which is it?

To find out, Mashable has been reviewing Anthropic's claims and talking to AI and cybersecurity experts.

What is Claude Mythos Preview?

Claude Mythos is a new large-language model that Anthropic says performs significantly better than Claude Opus 4.6, widely considered one of the best AI models in the world, especially in cybersecurity.

"In our testing, Claude Mythos Preview demonstrated a striking leap in cyber capabilities relative to prior models, including the ability to autonomously discover and exploit zero-day vulnerabilities in major operating systems and web browsers," reads the Claude Mythos system card.

Is Claude Mythos a sign of AGI?

Artificial general intelligence refers to superintelligent AI that can perform better than humans across a wide range of tasks. It's not an exaggeration to say that our entire economy has been organized around the quest for AGI, as Anthropic, Google, Meta, xAI, and OpenAI pour hundreds of billions of dollars into a new arms race.

If Claude Mythos is as capable as Anthropic says, would it be an example of AGI? The model card addresses this question directly, and Anthropic does seem to think it's close to AGI.

In a section about Claude Mythos safety risks, Antropic writes: "Current risks remain low. But we see warning signs that keeping them low could be a major challenge if capabilities continue advancing rapidly (e.g., to the point of strongly superhuman AI systems)." Of course, Anthropic has a strong financial incentive to promote this belief.

Ultimately, the model card for Claude Mythos is more conservative than the reaction online would suggest.

For example, while the Claude Mythos model card does show that this model performs above the trend line for previous Anthropic models, Anthropic says it does not show evidence of self-improvement or recursive growth. ("The gains we can identify are confidently attributable to human research, not AI assistance.")

Reasons to think Project Glasswing is a publicity stunt

Don't make me tap my sign: "[When] an AI salesman tells you that AI is an unstoppable world-changing technology on the order of the agricultural revolution...you should take this prediction for what it is: a sales pitch."

I wrote those words of caution in response to an essay by Anthropic CEO Dario Amodei that warned about the potentially cataclysmic dangers of AI. Anthropic also has a history of issuing dire warnings about its AI models.

You may remember the story of the Anthropic model that tried to "blackmail" a company CEO to prevent it from being turned off. In reality, Anthropic designed a test environment where blackmail was a potential outcome. This may be more akin to digital entrapment than genuine model misbehavior.

So, is Claude Mythos the latest example of the industry's Chicken Little problem?

On X, AI safety engineer Heidy Khlaaf listed a number of open questions that cast doubt on Anthropic's claims.

This Tweet is currently unavailable. It might be loading or has been removed.

Anthropic said the Claude Mythos preview found thousands of zero-day vulnerabilities. But Khlaaf says Anthropic left out key facts needed to assess this claim — the rate of false positives, how Claude Mythos compares to existing cybersecurity tools, and exactly how much manual human review was required.

"Releasing a marketing post with purposely vague language that clearly obscures evidence needed to substantiate Anthropic's claims brings into question if they are trying to garner further investment," Khlaaf told Mashable. "It also serves their 'safety first' image as they're able to frame the lack of public release, even a limited one for independent evaluation, as a public service when it simply obscures even experts' abilities to validate their claims."

We reached out to Anthropic repeatedly about these concerns, but the company did not respond. We will update this article if they do. In the Claude Mythos system card, Anthropic wrote that more data will be released in the coming weeks as the bugs Mythos found are patched and fixed.

Gary Marcus, an AI expert, author, and noted critic of the LLM hype machine, initially told Mashable that it was too soon to know whether Claude Mythos represented a new type of threat.

But Marcus has grown more skeptical since we spoke to him, and he recently wrote on X that Mythos was "nowhere near as scary" as it first seemed. "Folks, you can relax. Mythos is not some off-trend exponential gain," he wrote.

This Tweet is currently unavailable. It might be loading or has been removed.

Cybersecurity experts told Mashable it's also very unlikely Claude Mythos could be used to "turn off the lights" or bring down critical infrastructure.

"Claims about catastrophic uses of Mythos also significantly misunderstand threat models, cybersecurity risks, and the ability to propagate said risks in a way that could actually lead to safety-critical incidents," Khlaaf told us. "It's not as simple as asking a model 'hack this system,' with Anthropic's own technical blog post demonstrating a requisite of expertise that Anthropic downplays in their marketing posts."

Other experts expressed skepticism, while also acknowledging that Mythos does represent a genuine risk, which Marcus has also said.

"You could argue it didn’t need a public announcement," said Div Garg, a Stanford AI researcher and founder of AGI, Inc.
"However, ultimately, the decision to limit access to only those who develop and maintain critical software is precisely what you want a business to do in such a scenario…It’s easy to criticize the limited access, but worse outcomes would arise if they released it unchecked."

Tal Kollender, Founder and CEO of cybersecurity firm Remedio, told Mashable that tools like Claude Mythos are dangerous because they can exploit discovery.

"It's brilliant corporate theater," Kolender said. "Labeling a model 'too dangerous to release to the public' is certainly a marketing flex because it immediately creates mystique and signals immense power to investors. But beneath the PR stunt, there is a very real, very mundane truth...The cybersecurity industry doesn't actually have a 'finding' problem. We are already drowning in tools that detect vulnerabilities. What Mythos does is automate that discovery process at an unprecedented scale."

TL;DR: A week after revealing Claude Mythos Preview, some of Anthropic's biggest claims about the model look a lot sketchier, experts say. However, they also acknowledge that Claude Mythos, and other tools like it, pose a real risk.

Still, there are plenty of very valid reasons to be nervous about the new frontier model.

Reasons to think Claude Mythos Preview is a genuine threat to global cybersecurity

In the New York Times, author Thomas Friedman conjures a scenario straight out of War Games, where a teenager hacks the local power grid after school.

That scenario seems even more far-fetched a week later. But here's a much more likely scenario: A sophisticated group of hackers uses a tool like Claude Mythos to find zero-day vulnerabilities in our digital infrastructure, launching attacks faster than organizations can respond.

And that scenario should worry you.

If Claude Mythos isn't the tool that can do it, most experts agree such a tool isn't far off.

And some of the world's leading cybersecurity experts certainly seem worried.

"I've found more bugs in the last couple of weeks [with Claude Mythos] than in the rest of my entire life combined," said Nicholas Carlini, a research scientist affiliated with Anthropic and Google DeepMind, in a video on the Project Glasswing website.

"On Linux, we found a number of vulnerabilities where, as a user with no permissions, I can elevate myself to the administrator by just running some binary on my machine," Carlini said.

This week, the AI Security Institute published its findings on Claude Mythos's capabilities, and it provides some independent verification that it does represent a genuine leap forward.

Claude Mythos passed cybersecurity tests that no other model had ever completed, scoring higher than any other frontier model on virtually every test.

"Our testing shows that Mythos Preview can exploit systems with weak security posture, and it is likely that more models with these capabilities will be developed," AISI concluded.

This Tweet is currently unavailable. It might be loading or has been removed.

AISI also identified some limitations with Claude Mythos, which would impair its effectiveness in real-world scenarios.

So, was Anthropic’s rollout of Mythos responsible AI stewardship or self-serving marketing? Experts I talked to said these options aren’t mutually exclusive.

"I'd say it's both, and that's not a criticism," said Xu. "Any major platform rollout in this era is going to look different to different audiences depending on their fluency and their fear tolerance. What I care about is whether the intent is real, and the evidence I've seen from Anthropic suggests it mostly is."

As is often the case with fear-inducing AI headlines, the reality turned out to be more complicated.

"Personally, I don't go to bed worrying about a kid with Mythos hacking the power grid, but that doesn't mean the concern is fictional," said Howie Xu, Gen's Chief AI & Innovation Officer. "We're at an inflection point where the creative and collaborative upside of these tools is massive, and the security infrastructure hasn't caught up. That gap is exactly what keeps me busy. Even a fractional probability of a serious incident is too much, which is why building a trust and security layer into the agentic era is my extreme focus."

Finally, as Anthropic stresses in the Claude Mythos model card, tools like this will likely benefit cybersecurity defenders more than hackers in the long-term. And in the short-term, a more cautious approach — like the approach being modeled with Project Glasswing — may be warranted.

TL;DR: Claude Mythos has formidable cybersecurity coding abilities, and it does represent a genuine threat. However, if hackers have access to AI tools like Claude Mythos, so will the organizations defending against such attacks.