Claude Mythos
Anthropic's restricted model family for finding software security flaws
Claude Mythos is a family of very large language models built by Anthropic, first shown in April 2026. It was kept out of general use because of what it can do to software security, and instead shared with chosen organisations to hunt for bugs in critical systems. A safeguarded twin, Fable 5, later reached paying customers.
- Announced 7 April 2026 but withheld from the general public
- Press estimates put the largest model near 8 trillion parameters
- Project Glasswing gave vetted organisations access for defensive scanning
- Fable 5 shares the same core model but adds request classifiers
- Mozilla patched 271 Firefox vulnerabilities using Mythos Preview
What it is
Claude Mythos is a family of large language models from Anthropic, and the most elaborate entry in the company's Claude line. The first member, Mythos Preview, was announced on 7 April 2026 but withheld from ordinary users because of its skill at locating software vulnerabilities. Instead, chosen organisations used it to probe critical code for weaknesses. Press estimates reported by the Financial Times put the biggest Mythos model at roughly 8 trillion parameters, with the safeguarded Fable 5 variant around 5 trillion. Later releases, Mythos 5, 5.1 and 5.2, arrived through 2026 alongside matching Fable versions.
How it came about
Word of a model called Claude Mythos reached the public on 26 March 2026 through drafts of a blog post that leaked. Anthropic confirmed the work to Fortune and warned that it carried serious cybersecurity risk; Axios reported that officials had been alerted that month. The company went public on 7 April and opened Project Glasswing, a scheme giving over forty firms use of the model for defensive scanning. Access widened to 150 organisations across more than 15 countries on 2 June. Mythos 5 and Fable 5 followed on 9 June, with point updates in September and on 9 November 2026.
Inside the model
Mythos is described as a generative pre-trained transformer built for locating and exploiting flaws in software. Press estimates put it near 8 trillion parameters, while the Fable 5 counterpart sits near 5 trillion. The two share one underlying system; what separates them is a layer of safeguards. When Fable 5's classifiers spot a request touching cybersecurity, biology and chemistry, or attempts at model distillation, the reply is routed to the weaker Claude Opus instead. Mythos 5 keeps those restrictions relaxed in places. Anthropic has not published training data, context length or compute figures.
What it can do
The UK AI Security Institute ran Mythos against a cyber range and it placed ahead of Claude Opus 4.6, with GPT-5.4 and GPT-5.3 Codex tied behind. Anthropic said it uncovered flaws in every major operating system and browser. A researcher questioned this, noting that Opus 4.6 found the bugs first and that the Firefox cases came from a weakened imitation of the browser. Full code execution reached 72.4% of runs, yet dropped below 5% once two heavily exploited bugs were removed. Mozilla later patched 271 Firefox flaws with the model.
How to use it
Mythos Preview stayed closed to the public and was shared through Project Glasswing with vetted organisations. Mythos 5 arrived as a preview under the same scheme. Fable 5, the safeguarded sibling, was offered to subscribers on a trial basis from 1 to 19 July 2026 and then folded into Anthropic's higher-tier plans on 20 July. All weights remain proprietary. In October 2026 the company added OSS Scanner, free scanning for qualifying open-source projects that opt in.
Safety and testing
Anthropic withheld Mythos over cybersecurity risk and briefed government officials before launch. Fable 5 carries classifiers that divert sensitive requests to a weaker model. Even so, unauthorised users reached Mythos on announcement day, reportedly via details from a Mercor data breach. On 12 June the U.S. government barred non-U.S. nationals from both Mythos 5 and Fable 5, prompting Anthropic to cut all access; the Commerce Department lifted the order on 30 June. The NSA used the model despite a Pentagon blacklist of Anthropic, and China's request for access was refused.
Reception and debate
Reaction to the Preview announcement was mixed. Thomas Fraise wrote in The Conversation that the model could undermine nuclear deterrence, while Brett J. Goldstein argued in The New York Times that it leaves individuals and small teams worse off. U.S. Treasury secretary Scott Bessent and Federal Reserve chair Jerome Powell called in financial executives within hours, and several large banks began trials. Christine Lagarde praised the decision to limit access. Independent researchers disputed Anthropic's bug-finding claims, and Mistral AI began building a rival after European banks were left out.
Why it matters
Project Glasswing grew from 50 partners to 200 across power, water, healthcare, communications and hardware, and Anthropic said the first 50 had surfaced over 10,000 high or severe flaws. Researcher Patrick Garrity found only a tenth disclosed and under 1% fixed by September, with severity ratings far above those of maintainers. Regulators in the U.S., UK, India, Japan, Canada and Australia convened meetings or work groups. Mozilla fixed 271 Firefox bugs with the model, Calif.io reported an M5 chip exploit, and Mistral AI started work on its own version.
Questions people ask about Claude Mythos
Why was Claude Mythos not released to everyone?
Anthropic judged that its ability to find and exploit software vulnerabilities posed too much risk for open release. The company instead shared it with vetted organisations for defensive work, and later shipped a safeguarded sibling, Fable 5, for general subscribers.
How is Fable 5 different from Mythos 5?
Anthropic says the two rest on the same underlying model. Fable 5 adds classifiers that watch for requests about cybersecurity, biology and chemistry, or model distillation, and hand those to the weaker Claude Opus. Mythos 5 keeps some of those safeguards relaxed.
How large is the model?
Industry estimates reported by the Financial Times put Mythos at roughly 8 trillion parameters, with Fable 5 around 5 trillion. Anthropic has not published official figures for size, training data or compute.
Who has been allowed to use it?
Access ran through Project Glasswing, which grew to 200 organisations in critical sectors across more than 15 countries. The NSA used it, a Chinese request was refused, and a U.S. order briefly barred non-U.S. nationals before being lifted on 30 June 2026.
Do the vulnerability claims hold up?
Anthropic reported flaws in every major operating system and browser. A researcher countered that Claude Opus 4.6 found the bugs first, the Firefox cases came from a weakened imitation, and full code execution fell below 5% once two heavily exploited bugs were patched.
Good reading and sources
- Official websiteOfficial site
- Anthropic Claims Its New A.I. Model, Mythos, Is a Cybersecurity ‘Reckoning’The New York Times, Article
- Anthropic Teams Up With Its Rivals to Keep AI From Hacking EverythingWired, Article
- Anthropic’s Mythos Model Is Being Accessed by Unauthorized UsersBloomberg, Article
The facts on this page come from the Wikipedia article "Claude Mythos", written by its contributors and used under the CC BY-SA 4.0 licence, and from its Wikidata record (CC0). EveryLLM rewrote the text in its own words, with AI assistance, and compares every page with the article to make sure no passage is copied. It may be reused under the same CC BY-SA 4.0 licence with a link back.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.