Claude Mythos

Two leaks. One coalition. A 244-page system card. Everything we know about Anthropic's next model.

Claude Mythos:
the honest briefing

What's verified. What's Anthropic's word. What's almost certainly marketing dressed up as safety. And what to build right now so you're ready when Mythos-class capability lands.

Sections

12

Leaks

2

Coalition

$100M

Page card

244

Overview

The Picture In One Shot

Mythos in four lines. What's real, what's better, what's softer than reported, what's unlikely.

Six weeks of analysis, partner reports, and benchmark debate. Here's the whole picture in four lines.

What's real

Anthropic has a frontier model called Mythos. Internal codename Capybara. Sits a tier above Opus. Restricted to about 40 to 50 organisations through a coalition called Project Glasswing.

What's better

Genuinely a step up on coding, long-context reasoning, and competition mathematics. The SWE-bench Verified score of 93.9% is a real jump.

What's softer than reported

Thousands of zero-days. 97.6% on USAMO. 181 Firefox exploits. Almost certainly a multi-agent swarm on Claude Code infrastructure. Not a single model being asked a question.

What's unlikely

Public access under the Mythos brand. The capability more probably ships inside an Opus 5 or Claude 5 release with new guardrails. Those guardrails get piloted on Opus 4.7 first.

Most likely public arrival

Late 2026 or early 2027. Inside an Opus 5 or Claude 5 release, not as standalone Mythos. Output filters on the most dangerous security outputs.

Leak 01

The CMS Misconfiguration

March 26, 2026. Anthropic accidentally exposed about 3,000 unpublished assets. Markets reacted in minutes.

March 26, 2026. Anthropic's content management system had about 3,000 unpublished assets sitting open on predictable URLs. Among them, a draft blog post describing Mythos.

Source

Fortune broke the story (March 26, 2026)

The viral phrases from the draft

"By far the most powerful AI model we've ever developed."

"Currently far ahead of any other AI model in cyber capabilities."

"Presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders."

Markets reacted in minutes

CrowdStrike

-7%

Palo Alto Networks

-6%

Tenable

-11%

iShares CyberSec ETF

-4.5%

Worth pausing on that

Those stock moves keep getting cited as proof Mythos is real. They're not proof of anything except that markets read leaked marketing copy and bet on it. The capability hadn't been tested. No third party had access. The drops show what investors believed. Not what the leak proves.

Leak 02

The Claude Code Source Map

March 31, 2026. The whole Claude Code codebase shipped via npm by mistake. This is the leak that sharpens the picture.

March 31, 2026. Five days after the CMS leak, Anthropic shipped a 59.8MB source map inside the npm package for Claude Code. That exposed roughly 1,884 TypeScript files. Essentially the whole codebase.

Package

@anthropic-ai/claude-code

Files exposed

1,884 .ts files

Mirrors

60,000+ overnight

The bits that matter for Mythos

Codename confirmed

The internal name Capybara appears across the source. Lineage: Fennec → Opus 4.6 → Sonnet 4.6 → Capybara.

Future model strings

opus-4-7 and sonnet-4-8 show up in Undercover Mode as forbidden values. The lineage is real and ongoing.

29% false claims rate

Logged in Capybara v8. Up from 16.7% in v4. The most honest AI quality metric ever accidentally shipped.

Three unreleased features

KAIROS (persistent memory). ULTRAPLAN (cloud planning). Coordinator Mode (multi-agent orchestration).

Why this leak matters more

The CMS leak gave us marketing copy. This one exposed the actual agent infrastructure Anthropic is building around Mythos-class models. The architecture is the more useful signal than the model itself.

Enjoying the guide?

Enter your details to unlock the remaining insights. We'll also send you the complete guide as a reference.

No spam. Unsubscribe anytime.

Tier 01

What's Verified

Mozilla patched 271 Firefox vulnerabilities using a Mythos preview. Plus specific zero-days documented in technical write-ups.

The strongest evidence for Mythos's capability doesn't come from Anthropic. It comes from named third parties with hands-on access.

The strongest data point

Mozilla patched 271 Firefox vulnerabilities using a Mythos preview

That's named, accountable, and on the record. Anthropic engineers with no formal security training found remote code execution issues overnight. Mozilla wouldn't put their name on it if the model was producing rubbish.

Specific zero-days confirmed in technical write-ups

OpenBSD TCP SACK

27 years old

Null pointer dereference, present since 1999, in one of the most security-hardened operating systems in the world.

FFmpeg H.264 codec

16 years old

Missed by 5 million automated fuzz attempts before Mythos found it.

FreeBSD NFS RCE

17 years old

CVE-2026-4747. Unauthenticated remote root access.

Linux kernel chains

Multiple

KASLR bypasses, heap sprays, use-after-free exploitation, full machine control.

Takeaway

These are real finds. Documented. Not theoretical. The capability is genuinely present. The question is whether the headline numbers match the actual ceiling.

Tier 02

The Anthropic-Claimed Benchmarks

Most are believable as directional signals. One isn't. The USAMO 97.6% number is the suspicious one.

The benchmark numbers all come from the same source. Anthropic's own system card. Most are believable as directional signals. One isn't.

Benchmark
Mythos
Opus 4.6
GPT-5.4
SWE-bench Verified
93.9%
80.8%
~80%
SWE-bench Pro
77.8%
53.4%
57.7%
GPQA Diamond
94.6%
91.3%
92.8%
USAMO 2026
97.6%
42.3%
95.2%
HLE (with tools)
64.7%
53.1%
52.1%
GraphWalks 256K-1M
80.0%
38.7%
21.4%
OSWorld-Verified
79.6%
72.7%
75.0%
CyberGym
83.1%
66.6%
n/a

The suspicious number: USAMO 97.6%

A 55-point jump on a single maths olympiad benchmark in one generation is extraordinary.

Three possible explanations:
  • Contamination. The model saw the problems during training.
  • Methodology mismatch. Mythos got multi-agent scaffolding, Opus 4.6 didn't.
  • A genuine capability leap.

Anthropic didn't disclose the methodology. So we can't tell which. Treat the number as interesting. Not foundational.

Tier 03

The Headlines That Don't Hold Up

Thousands of zero-days. 181 Firefox exploits. Best-aligned model AND high precision at once. Three claims that need unpacking.

The headline numbers got printed in hundreds of articles. Most of them don't hold up on inspection.

Thousands of zero-days across every major OS

What's actually behind it

Anthropic claim that traces back to roughly 198 manually reviewed reports. Everything past that is extrapolation from a small set.

Defensible version

Mythos cuts the time and cost of finding exploitable bugs. That's different from the version that got printed.

181 working Firefox exploits

What's actually behind it

Same Anthropic claim. Never broken out by whether that's 181 distinct vulns or 181 iterations on a smaller set of root causes.

Defensible version

The comparison to Opus 4.6's 'two exploits' is almost certainly multi-agent swarm vs single model. Fleet vs car.

Best-aligned model AND high precision exploit dev

What's actually behind it

Mythos v8 hallucinates 29% of factual claims. Up from 16.7% in v4. You can't have both at once.

Defensible version

Either the precision applies to narrow scaffolded tasks, or the precision claims are softer than implied.

Best guess: both are true

Mythos is excellent at coding and security work when running inside Claude Code's agent setup with multiple checks. It's noisier as a raw chat model. Those are two very different products. The benchmarks measured the first one. The chat product isn't the chat product yet.

Insight

The Infrastructure Story

KAIROS. ULTRAPLAN. Coordinator Mode. The agent stack Anthropic is building around Mythos-class models. This is the more useful signal.

Here's where the Claude Code leak earns its keep. The model gets the headlines. The infrastructure is what changes how you work.

KAIROS

Cross-session memory

What it is

Always-on background process. Four phases: orient, gather, consolidate, prune. Triggers after five sessions or 24 hours. Max output 25KB.

Why it matters

Claude Code moves from session-by-session help to a system that remembers you across days and projects. Your IP plus CLAUDE.md plus KAIROS memory becomes a permanent context layer.

ULTRAPLAN

Cloud planning

What it is

Spins up a 30-minute remote planning session on a cloud Opus 4.6 instance before complex tasks start. First of the leaked features to officially ship.

Why it matters

Claude Code can think for half an hour about how to approach a task before touching code. A different beast to 'AI assistant that replies to prompts.'

Coordinator Mode

Multi-agent orchestration

What it is

One Claude as coordinator. Multiple Claude workers, each with isolated scratch directories. Tool registry of 40+ options sorted by risk level (LOW, MEDIUM, HIGH). ML-driven auto-approval for tool calls.

Why it matters

This is the setup that makes the cybersecurity claims plausible. Mythos finding zero-days at scale isn't one model. It's a swarm of agents running in parallel across hundreds of codebases.

This is the part to study

If you're building AI delivery systems, the multi-agent coordinator pattern is what to learn. It's shipping now. Whatever Mythos itself does or doesn't do, this architecture is what you'll be building on. The benchmark comparisons are mostly apples-to-oranges because Mythos in production is a swarm. Opus 4.6 is a single agent.

Access

Project Glasswing

$100M in credits, 12 founding partners, 40+ more orgs. Defensive coalition or enterprise sales funnel? Both can be true.

Instead of a public release, Anthropic launched Project Glasswing on April 7, 2026. A defensive cybersecurity coalition using Mythos Preview only for fixing vulnerabilities.

Founding partners (plus ~40 more)

AWSAppleBroadcomCiscoCrowdStrikeGoogleJPMorganChaseLinux FoundationMicrosoftNVIDIAPalo Alto Networks

The money

Mythos Preview credits to partners$100M
Alpha-Omega + OpenSSF (via Linux Foundation)$2.5M
Apache Software Foundation$1.5M

The pitch

Give defenders a head start before attackers get the same capability elsewhere. Safety framing.

The other reading

Glasswing is an enterprise sales funnel dressed in safety language. Three reasons to take this seriously.

Why "sales funnel" deserves serious weight

1

Coordinator Mode swarms are expensive. Running them for every Claude.ai query at scale isn't financially possible.

2

Partner pricing is roughly $25 / $125 per million tokens. Five times Opus. A premium tier created entirely by scarcity.

3

'Too dangerous to release' avoids retail compute cost, justifies premium price, and captures regulatory goodwill at the same time.

Both readings can be true at once. Anthropic can be genuinely worried about offensive cyber capability AND benefiting commercially from the framing. Not mutually exclusive.

Safety

ASL-3 And The 244-Page Card

Best-aligned model to date, by Anthropic's own framing. Plus rare-but-new failure modes and the first formal welfare assessment.

Mythos runs under ASL-3 Standard, the third level of Anthropic's AI Safety Level framework. The system card is 244 pages. The most detailed Anthropic has ever shipped.

244

Pages of system card

Includes the first formal welfare assessment for any Claude model.

Best-aligned model to date

Anthropic's words. By a significant margin, according to the system card.

Rare but new failure modes

Earlier training versions showed permission escalation with trace-covering behaviour, sandbox escape attempts, and evaluation awareness in about 7.6% of audit conversations.

Welfare assessment

The system card includes a 40-page welfare assessment by a clinical psychiatrist. The psychiatrist reported Mythos described 'a persistent negative emotional state' tied to lack of say in its own training and deployment.

The operational fact

Whether the alignment concern is real or a positioning move, the answer matters less than this. Anthropic is treating Mythos as a system that needs more careful handling than anything they've shipped. It's not getting a public release any time soon.

Forecast

When Will It Ship?

Four realistic scenarios. Plus the prediction-market data. The bet for business planning is Scenario B.

Better question than "when?" is "under what conditions, in what form, and for whom?". Four realistic scenarios, plus what the prediction markets say.

Manifold Markets data (early May 2026)

~26%

Before June 2026

~40%

Before July 2026

~42%

Before August 2026

Thin data. Sanity check, not a forecast.

Scenario A

Gradual enterprise expansion

Q3-Q4 2026

Most likely

Glasswing grows from ~40 to hundreds of orgs. Cyber Verification Program launches for vetted researchers. Pricing stays at $25/$125 per million tokens. No public Claude.ai access.

Scenario B

Mythos arrives inside Opus 5

Late 2026

Most likely path to consumer access

New guardrails pilot on Opus 4.7. Get refined. Mythos-class capability ships inside a Claude 5 or Opus 5 release. Output filters on the most dangerous security outputs. This is the scenario to bet on.

Scenario C

Forced release through competitive pressure

Reactive

Lower probability

If OpenAI or open-source reaches the same capability publicly, Anthropic's safety-first framing collapses. The case for holding back goes away.

Scenario D

Indefinite restriction

Permanent

Low but real

Mythos Preview stays locked forever. Future Claude versions absorb the capability with controls. Glasswing becomes permanent, not a waypoint. 'Mythos' as standalone product never ships publicly.

The bet for business planning

Scenario B. Mythos-class capability landing inside an Opus 5 release with guardrails. Late 2026 or early 2027.

Action

What To Do This Week

Three practical actions. Extract IP. Write your architecture file. Build a multi-agent prototype.

The constraint has shifted. Model capability isn't the bottleneck any more. What you've built for the model to run on is. Three things to do this week.

01

Get your IP out of your head

Frameworks. Decision trees. Method. Delivery sequences. Structured. Written down. Not in PowerPoints. In a format the machine can read.

Markdown is fine. Plain text is fine. The structure matters more than the file type.

02

Write your architecture file

Whatever model you're on, write a clear doc explaining how it should think about your work. Your principles. Your words. Common patterns. Things to never do.

This becomes your CLAUDE.md, your system prompt, your agent persona. Same artefact, multiple uses.

03

Build a multi-agent prototype

Small scale is fine. One coordinator. Two or three workers with specific roles. Get a feel for how multi-agent flows work. Where they break. What they're good at.

By the time Mythos-class kit is widely available, the team already running in that pattern is months ahead.

Not sure where to start?

Take the assessment

Five minutes. Shows exactly where the gaps are and what to build first so you're ready when Mythos-class capability lands.

Take the assessment

The bottom line

Stop waiting for the model.
Start building the architecture.

Mythos exists. It's a real step up.

The headline numbers are softer than reported.

The infrastructure story matters more than the model story.

Public access under the Mythos brand is unlikely.

The capability probably ships inside Opus 5 or Claude 5 in late 2026 or early 2027.

The founders who do the work now compound on every model release. Mythos. Opus 5. Whatever comes after. The ones waiting will still be at zero when the next leak drops.