Claude Mythos Preview, explained: capabilities, access, and cybersecurity risks

That matters because the name has invited speculation. Available sources do not support treating Claude Mythos Preview as a separate company, a standalone consumer app, or a regular Claude tier that anyone can freely select.^[2][5][8]

What Anthropic says about its capabilities

The clearest capability claim comes from Anthropic’s system-card material: Claude Mythos Preview is significantly more capable than Claude Opus 4.6.^[8] The same material describes Opus 4.6 as the most capable model covered in Anthropic’s most recent risk report, making Opus the main comparison point for Mythos.^[8]

The nuance is important. Anthropic also says Mythos Preview’s profile is effectively similar for the purposes of its overall risk assessment, and that some multi-turn evaluation results were comparable to Claude Opus 4.6 and Claude Sonnet 4.6 within the margin of error.^[8]

Why cybersecurity dominates the Mythos story

The most concrete public technical details about Mythos involve vulnerability and exploit-related testing. Anthropic’s red-team write-up says Mythos Preview was able to write sophisticated exploits fully autonomously for already-patched bugs.^[4] It also says the model fully autonomously identified and exploited a 17-year-old remote-code-execution vulnerability in FreeBSD that could allow root access on a machine running NFS.^[4]

Those findings do not mean Mythos is publicly available as an offensive cybersecurity tool. They do show why access control, monitoring, red-teaming, and abuse prevention are central to the model’s rollout. TechCrunch frames its coverage of Mythos around security, and CrowdStrike’s coverage makes the same broader point: more capable AI systems require stronger security practices.^[1][2]

Is Claude Mythos Preview available to use?

The available sources do not show Claude Mythos Preview as a regular, broadly available Claude option. TechCrunch describes Mythos as a preview, and Fortune reported that Anthropic had begun testing the model with early-access customers after a data leak revealed its existence.^[2][5]

For developers and security teams, the practical takeaway is to avoid planning around Mythos as a generally deployable model until Anthropic publishes official access and deployment guidance. The current public evidence is strongest on model positioning, comparative capability, and security testing—not broad availability.^[2][4][5][8]

How to read the leak reporting

Fortune reported that draft material in an unsecured, publicly searchable data store revealed the model’s existence, described it as Claude Mythos, and said Anthropic believed it posed unprecedented cybersecurity risks.^[5] That is an important part of the story, but it should be read carefully because it came from reported draft material rather than a final public product page.^[5]

Anthropic’s own system-card material is more measured: it says Mythos Preview is more capable than previous models, while also saying its profile is effectively similar for the purposes of the overall risk assessment.^[8] The best reading is not that the sources cleanly contradict each other, but that Mythos combines a major capability claim with especially sensitive cybersecurity behavior in testing.^[4][5][8]

What not to assume

Mythos is not just a renamed Opus

Anthropic directly compares Claude Mythos Preview with Claude Opus 4.6 and says Mythos is significantly more capable.^[8] That makes a simple rebrand explanation a poor fit for the available evidence.

The capability claim is not a blank check

Anthropic’s strongest claim is about overall capability versus Opus 4.6.^[8] The same system-card material says some disclosed multi-turn evaluation results were comparable to Opus 4.6 and Sonnet 4.6 within the margin of error, so the evidence should not be stretched into a claim that Mythos dominates every task.^[8]

A security-heavy preview is not the same as public exploit access

Anthropic’s red-team findings are significant because they describe autonomous exploit-related behavior in testing.^[4] They do not, by themselves, show that the model has been released for unrestricted offensive use or broad public access.^[2][4][5]

FAQ

Is Claude Mythos the same as Claude Opus?

No. Anthropic compares Claude Mythos Preview with Claude Opus 4.6 and says Mythos Preview is significantly more capable.^[8]

Is Claude Mythos a cybersecurity model?

The sources identify Claude Mythos Preview as a Claude model preview, not as a standalone cybersecurity product.^[2][8] Cybersecurity is central to the public discussion because Anthropic’s red-team material describes autonomous exploit-writing and vulnerability exploitation during testing.^[4]

Can anyone use Claude Mythos Preview?

The cited public sources do not show Mythos as a normal public Claude release. TechCrunch describes it as a preview, while Fortune reported early-access customer testing after a data leak revealed the model’s existence.^[2][5]

Bottom line

Claude Mythos Preview matters because it pairs Anthropic’s claim of a major capability step beyond Claude Opus 4.6 with concrete red-team evidence of advanced autonomous vulnerability exploitation.^[4][8] The prudent read is that Mythos is an Anthropic Claude model preview, apparently not a broad public launch, and its most consequential story is security governance rather than model rankings alone.^[2][4][5][8]