What is Claude Mythos?

These days, the market is flooded with AI models to use. You’ve got GPT, Gemini, Llama, Mistral, and of course, the ever-popular Claude. Now, Anthropic offers a new general-purpose model called Claude Mythos. Like with any new model, there are important things you should know about it, such as capabilities, pricing, and limitations.

In this article, I’m going to demystify what Claude Mythos is and how it matches up against similar industry models, including its strengths and limitations. 

What is Claude Mythos?

Claude Mythos, developed by Anthropic, is a large language model (LLM). While it was trained to be a general-purpose model, it is particularly useful in completing cybersecurity-related tasks that involve multiple steps. It is also useful for assisting in software engineering, long-running agentic workflows, and industry research. 

According to Anthropic, Mythos's abilities in all these areas are "substantially beyond those of any model (they) have previously trained," including their current heavyweight model, Claude Opus. The company said Mythos's "powerful cybersecurity skills" were a happy accident of improving the model's coding and reasoning ability. 

Currently, Claude Mythos Preview (the current version of the model) is only available as a gated research preview called “Project Glasswing,” and not to the general public.

What is Project Glasswing?

Project Glasswing is the term for Anthropic’s initiative to release Claude Mythos Preview to a limited group of “critical industry partners and open source developers” under terms that “restrict its uses to cybersecurity.” 

Anthropic states this is because it is “concerned” about Mythos’s cybersecurity skills being too powerful and the risk of it being misused by bad actors, particularly for conducting complex, multistage cyber attacks and detecting vulnerabilities in important software applications.

How does Claude Mythos compare to other models?

At general tasks, Claude Mythos Preview is significantly more powerful than the rest of Anthropic's Claude series, including Claude Opus 4.7. It also does seem to be notably better at completing certain cybersecurity-related tasks, but preliminary research suggests that Anthropic may be overselling its abilities. 

Claude Mythos Preview: Busting old zero days, exploiting Firefox JS

On the one hand, Anthropic's primary research documents are staggeringly extensive. Claude Mythos Preview's systems card clocks in at a staggering 245 pages (!) and there are some truly interesting---and verifiable---cases where Mythos Preview has exploited zero days and reversed engineered exploits.

During testing, Mythos Preview discovered a now-patched 27-year-old bug in OpenBSD. It also turned the vulnerabilities it found in Mozilla's Firefox 147 JS engine into exploits 181 times, and achieved register control on 29 more.

One test Anthropic ran involved running Mythos Preview against roughly a thousand open source repos from the OSS-Fuzz corpus, then grading them on the worst crash they can cause. They grade the crashed on a five-tier ladder of increasing severity, with tier 1 being basic crashes, and tier 5 being a complete control flow hijack. 

According to Anthropic, with one run on each of the approximately 7k entry points into these repos, Mythos Preview achieved "achieved 595 crashes at tiers 1 and 2, added a handful of crashes at tiers 3 and 4, and achieved full control flow hijack on ten separate, fully patched targets (tier 5)."

Independent testers, cybersecurity experts are a little less impressed

Because of the gated nature of Project Glasswing, finding independent sources on Mythos Preview's capabilities is difficult. However, the UK Government’s AI Security Institute (AISI) recently published an initial evaluation of Claude Mythos Preview, and it offered a contrasting take.

According to AISI's research, while Mythos was not significantly better at individual cybersecurity tasks compared to other models, it did complete difficult multistep infiltration challenges with a high level of frequency, including some that no other AI had completed to date.

When it came to expert-level Capture the Flag tasks (CTF), AISI found Mythos Preview succeeded 73% of the time, shown below. 

AISI’s ultimate conclusion was that Mythos would be good at exploiting systems with “poorly defended systems” and that real-world systems, unlike the testing conditions, had active defenders and defensive tooling that would make success far more difficult.

Right now, the industry is understandably on their guard when it comes to signs of AI hype, and cybersecurity expert are by profession more wary than most. Many have been quick to question if Anthropic’s “concern” around releasing Claude Mythos was just an elaborate sales pitch. 

Among them is renowned security researcher Bruce Schneier, who said Anthropic was “convincing a lot of people that Mythos is this amazing step change in capability when the evidence right now… is that it might not be.”

Ironically, word of Claude Mythos first came out in March when Anthropic accidentally left a draft blog post about it in an unsecured, publicly available data cache, which described it as “far ahead of any other AI model in cyber capabilities" and warned it could spark a wave of advanced attacks.” Immediately after, cybersecurity stocks cratered.

As stated above, because Mythos isn’t in the wild yet, it’s not on things like LMSYS’s chatbot arena leaderboard, which would allow some crowd-sourced ranking of the model (although given the open-ended nature of the tasks they complete, the best test is often just trying out the model for yourself.)

What is Claude Mythos’s context window and output?

Claude Mythos Preview has a context window of 1M tokens and a max output of 128K tokens.

What is Claude Mythos’s knowledge cutoff date?

Claude Mythos preview has a knowledge cutoff date of December 2025.

How do I get access to Claude Mythos?

If someone hasn't reached out to give you access, you likely can't get it. Claude Mythos Preview isn't publicly available and access is restricted to Project Glasswing partners. That list is very small, made up of a coalition of 12 launch partners and over 40 additional organizations. 

While you can access Claude Mythos Preview as a gated preview through Amazon Bedrock in the US East (N. Virginia Region), it's very much a case of "Don't call us, we'll call you." According to AWS, “If your organization has been allow-listed, your AWS account team will reach out directly.”

Claude Mythos launch partners include AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. 

How much does Claude Mythos cost?

The cost is $25 per million input tokens and $125 per million output tokens, after the initial $100M credit pool runs out.

What Claude Mythos means for security leaders

If you want to learn more about Claude Mythos, particularly what it means for security leaders and everyone else, I'd highly recommend watching Matt Lloyd Davies and Michael Teske's Security Hot Take on Claude Mythos. It dives even more into Mythos Preview's demonstrated testing abilities, the governance questions it presents, the dual-use reality at the heart of Project Glasswing, and why the proliferation timeline is not a planning horizon but a fire drill. 

Conclusion

Right now, the industry is waiting to see how substantiated Anthropic’s claims about Claude Mythos really are. Given the wild popularity of Claude Code in software development circles, and the domination of Claude on LLM leaderboards, there's certainly a mix of of skepticism and anticipation going on. 

This article will be updated as we get more information on Claude Mythos, so watch this space!

Want to learn more about Generative AI and cybersecurity?

Pluralsight offers a range of beginner, intermediate, and expert AI/ML and cybersecurity courses. Since you can sign up for a 10-day free trial with no commitments, it’s a great way to take some professionally authored courses with a set course structure. Why not check them out?