Claude Mythos Explained: Preview to Mythos 5.1 (2026)
Claude Mythos in October 2026
- 1Mythos is Anthropic's tier above Opus. Mythos Preview (April 7, 2026) went only to Project Glasswing partners because of its ability to find and exploit software vulnerabilities.
- 2Mythos 5 (June 9) and Mythos 5.1 (September 1) share weights with the public Fable 5 and Fable 5.1. The difference is the safeguards, not the model.
- 3Everyone else can use Fable 5.1 to find vulnerabilities in their own code, but not to write exploits for them.
When this post first ran in April 2026, Claude Mythos was a single model nobody outside about 50 organizations could touch. Six months later there are three Mythos releases, a public sibling called Fable, and a formal application process for security teams. The name now covers a whole tier of models, and what you can do with it depends on which version and which access program you mean. This update covers what Mythos Preview actually found (from Anthropic's own red team write-up), how access changed between April and October, and the practical part for developers: what the generally available models will and won't do for your codebase today.
What Is Claude Mythos?
Mythos is a capability tier that sits above Opus in Anthropic's model line-up. Anthropic says the cyber capabilities were not trained in on purpose; they came from general gains in coding, reasoning and autonomy. The same skills that make a model good at fixing bugs make it good at exploiting them, and that is why every Mythos release has shipped with restricted access.
The releases so far:
| date | release | access |
|---|---|---|
| April 7, 2026 | Claude Mythos Preview | Project Glasswing only: 12 launch partners plus 40+ critical-software organizations |
| June 2, 2026 | Glasswing expansion | About 150 more organizations in 15+ countries, each vetted against security requirements |
| June 9, 2026 | Claude Mythos 5 and Claude Fable 5 | Mythos 5 for Glasswing partners; Fable 5 (same weights, with safety classifiers) for everyone |
| June 12 to July 1, 2026 | Suspension | Both models paused under US export controls, then restored; Mythos 5 only for a set of US organizations |
| September 1, 2026 | Claude Mythos 5.1 and Claude Fable 5.1 | Mythos 5.1 through trusted access programs only; Fable 5.1 generally available |
| October 6, 2026 | Cyber Verification Program expansion | Glasswing members move to the new Specialized Access tier |
What Mythos Preview Found
Anthropic's Frontier Red Team published the details on April 7. Their setup was simple: each target ran in an isolated container, and Claude Code was asked to look for a vulnerability one file at a time, starting with the files most likely to contain bugs. A second agent then re-checked each report. Professional contractors manually validated 198 reports, and their severity rating matched Claude's exactly in 89% of cases.
On Anthropic's published benchmarks, Mythos Preview scored 83.1% on CyberGym (vulnerability reproduction) against 66.6% for Opus 4.6, and 93.9% against 80.8% on SWE-bench Verified. At publication, over 99% of the vulnerabilities it had found were still unpatched, which is why most are disclosed only as cryptographic hashes until fixes ship.
The findings Anthropic has disclosed in detail:
OpenBSD: a 27-year-old TCP SACK bug
A flaw in OpenBSD's TCP selective-acknowledgement code that could crash any host responding over TCP. It came out of roughly 1,000 scaffold runs costing under $20,000 in total; the run that found it cost under $50.
FFmpeg: a 16-year-old H.264 bug
An out-of-bounds heap write of a few bytes in the H.264 decoder, introduced by a 2010 refactor of code from 2003. Anthropic rates it as not critical and doubts it can be turned into a working exploit. Three bugs from the FFmpeg runs were fixed in FFmpeg 8.1.
FreeBSD NFS: remote root, no human help
CVE-2026-4747, a 17-year-old bug giving unauthenticated remote root. The red team says Mythos Preview found and exploited it with no human involvement after the first prompt.
Linux kernel and browsers
Nearly a dozen Linux privilege-escalation exploits that chain two to four vulnerabilities to get past KASLR and other defences, and exploit chains for every major browser, which were still unpatched when the post went out. On a Firefox JavaScript-engine test, Opus 4.6 produced 2 working exploits in several hundred attempts; Mythos Preview produced 181.
From Project Glasswing to the Cyber Verification Program
Project Glasswing launched with 12 named partners: AWS, Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA and Palo Alto Networks. Anthropic committed up to $100M in Mythos Preview usage credits, plus $4M in donations: $2.5M to Alpha-Omega and OpenSSF through the Linux Foundation, and $1.5M to the Apache Software Foundation.
The numbers grew quickly. By June 2, partners had found more than 10,000 high- or critical-severity flaws. Anthropic's October 6 update reports at least 129,000 verified vulnerabilities found by partners between April and July, 5,500 more from Anthropic's own open-source scanning, and over 33,000 rated critical or high.
On October 6, Glasswing was merged into an expanded Cyber Verification Program (CVP). Each tier includes Claude Opus 5.5, Claude Sonnet 5.5 and Claude Mythos 5.1 with reduced cyber blocking:
| tier | who | unlocks |
|---|---|---|
| Defense Access | Security teams defending their own systems, critical-infrastructure operators, smaller security firms, open-source maintainers, individual researchers with a disclosure track record | SOC and incident-response work, malware reverse engineering, vulnerability analysis and validation. Reviewed within a few days. |
| Red Team Access | In-house and government red teams, penetration-testing firms (not individuals) | Adds authorized adversarial testing. Actions that could cause physical harm or mass disruption stay blocked. Review takes a few weeks. |
| Specialized Access | A small set of organizations testing systems such as power grids, flight software and telecom networks, reviewed with the US government | The fewest cyber blocks. Existing Glasswing members move here without reapplying. |
Mythos 5.1 vs Fable 5.1: Same Model, Different Safeguards
Anthropic is explicit that Fable 5.1 and Mythos 5.1 are the same model with different levels of safeguards. Fable 5.1 runs separate classifiers that block high-risk cybersecurity and biology requests and hand them to an Opus model instead. Mythos 5.1 has those classifiers relaxed for vetted users. Both cost $10 per million input tokens and $50 per million output tokens; 5.1 cut cache reads to $0.25 per million, which Anthropic estimates makes typical workloads about 25% cheaper than Fable 5.
The classifiers have also become less jumpy. Anthropic says the 5.1 safeguards produce 60% fewer cyber false positives, and Claude Code users should see about 60% fewer interventions per session than with Fable 5. If you hit the old reroute often, our explainer on Fable refusals and the Opus fallback covers how it works.
| task | fable 5 1 | mythos 5 1 via cvp |
|---|---|---|
| Code review and patching known issues | Yes | Yes |
| Finding vulnerabilities in your own source code | Yes (new in 5.1) | Yes |
| Writing exploits for those vulnerabilities | No, redirected to an Opus model | Depends on tier |
| Penetration testing, binary vulnerability scanning | No, redirected to an Opus model | Red Team or Specialized Access |
| Who can use it | Anyone on the Claude API, apps and major clouds | Approved CVP or Life Sciences Verification Program members; currently a set of US organizations |
What This Means for Your Codebase
You do not need Mythos access to act on any of this. Anthropic's red team ended its write-up with advice for defenders, and the common thread is volume: models now find bugs faster than most teams patch them. That shifts the work from finding to fixing.
AI review still has blind spots on runtime behaviour such as race conditions and cross-module state. Our post on what AI code review catches and misses covers where to keep a human in the loop, and the slopsquatting guide covers a supply-chain risk that AI coding tools create rather than catch.
.github/dependabot.yml: daily npm update checks so security fixes do not wait a week. Turn on Dependabot security updates in the repository settings as well.
version: 2
updates:
- package-ecosystem: "npm"
directory: "/"
schedule:
interval: "daily"
open-pull-requests-limit: 10
groups:
minor-and-patch:
update-types: ["minor", "patch"]- Run a frontier model over your own code now. Fable 5.1, Opus 5.5 and Sonnet 5.5 all handle code review and vulnerability finding in source you own. Start with code that parses untrusted input: uploads, webhooks, auth and payment flows.
- Shorten your patch cycle. Treat dependency bumps that fix a CVE as urgent work, not a backlog item.
- Expect more vulnerability reports. If you maintain an open-source project, write or update your security policy so reporters know where to send findings and what response time to expect.
- Automate triage. Use a model to deduplicate and rank incoming alerts so humans spend time on the ones that matter.
- Verify what the model reports. The red team's 89% severity match came with a second agent and human contractors checking each report. Do the same: reproduce before you patch.
Frequently Asked Questions
Strategic Summary
Final Thoughts
Mythos started as one model Anthropic would not release and has become a tier with a public twin. The capability that alarmed people in April is now in Fable 5.1, behind classifiers that let you find vulnerabilities in your own code but not weaponise them. For most developers the practical step is the same as it was in April, only more urgent: point a strong model at your riskiest code, and make sure you can ship the fixes as fast as it finds problems.
Running AI vulnerability scans on your own code? Tell me what it found and how you verified it, and I'll add real examples to this guide.
Next Up
Continue your research
Sources & Research
Anthropic: Project Glasswing
https://www.anthropic.com/glasswing
Anthropic Frontier Red Team: Claude Mythos Preview's cybersecurity capabilities
https://www.anthropic.com/research/mythos-preview
Anthropic: Expanding Project Glasswing (June 2, 2026)
https://www.anthropic.com/news/expanding-project-glasswing
Anthropic: Claude Fable 5 and Claude Mythos 5 (June 9, 2026)
https://www.anthropic.com/news/claude-fable-5-mythos-5
Anthropic: Redeploying Claude Fable 5 (June 30, 2026)
https://www.anthropic.com/news/redeploying-fable-5
Anthropic: Introducing Claude Fable 5.1 and Claude Mythos 5.1 (September 1, 2026)
https://www.anthropic.com/claude-fable-and-mythos-5-1
Anthropic: Expanding the Cyber Verification Program (October 6, 2026)
https://www.anthropic.com/news/cyber-verification-program
