Claude Mythos Preview (Anthropic’s unreleased frontier model) now has a detailed public…

Market Intelligence Summary

Query: Mythos Update

Generated: May 25, 2026 at 09:45 PM

Executive Summary

Claude Mythos Preview (Anthropic’s unreleased frontier model) now has a detailed public write‑up via the Project Glasswing “initial update”, showing it can autonomously find and exploit vulnerabilities at a level comparable to highly skilled professional security researchers.[1][7]
In early Glasswing runs, Mythos Preview identified over 10,000 vulnerabilities, including 6,202 that it internally rates as high or critical severity across both open‑source and proprietary “globally important” software.[1][7]
Despite this, patching is badly lagging: fewer than ~1% of vulnerabilities found by Mythos have been patched so far; on open source, only 75 of 530 high/critical bugs reported were fixed at the time of the report, with a typical patch latency of around two weeks.[1][5]
The update confirms Mythos will not be generally released in its current preview form due to its offensive cyber capabilities; instead, access is being gated via Project Glasswing and $100m in free credits for vetted partners to use it defensively.[1][4][7]
External analysts (e.g., Schneier, Cloudflare, independent security blogs) view Glasswing as both a real technical step change and a highly staged, capital‑intensive “demo environment” that exposes a structural problem: AI dramatically accelerates bug-finding, but the remediation pipeline and incentive structure have not caught up.[2][5][8][9]

---

Comprehensive Analysis

Context: What Mythos and Project Glasswing Are

Anthropic has developed Claude Mythos Preview, a “frontier” large language model trained primarily to be extremely strong at code, with cybersecurity capabilities emerging as a side‑effect rather than from explicit offensive training.[6][7] According to Anthropic’s own Glasswing write‑up and system card, early internal evaluations indicated Mythos was meaningfully better than prior models (e.g., Claude Opus 4.6 and Sonnet 4.6) at both finding and exploiting software vulnerabilities, including chaining multiple bugs into complex exploits.[4][5][7]

In response, Anthropic did two things:

Declared that Mythos Preview would not be released to the general public, citing the risk that the same capabilities used to secure systems could be weaponized by attackers.[2][4][7]
Launched Project Glasswing, a coordinated initiative with major technology and infrastructure players (e.g., Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, Palo Alto Networks, etc.) to run Mythos against critical software stacks and patch the resulting findings before adversaries obtain comparable models.[6][7][8]

The latest “initial update” on Glasswing and Mythos, which your query references, pulls these strands together and provides quantitative detail on Mythos’ observed performance and the emerging remediation bottleneck.[1][7]

---

Mythos Update: Core Technical and Operational Findings

Scale and severity of vulnerabilities discovered

From the Anthropic Glasswing update and cross‑summaries:

Total vulnerabilities found: Mythos Preview has discovered more than 10,000 vulnerabilities across the codebases it has been run against.[3][7][10]
High/critical subset: Within that, 6,202 vulnerabilities are rated by Anthropic’s internal scoring as high or critical severity.[1][5][7]
Open‑source share: In open‑source projects alone, Mythos has found 23,019 vulnerabilities in total, with 6,202 classified as high/critical.[1]
Patch rate: For open source, only 75 of 530 high/critical bugs reported were patched at the time of the update, and 65 of those 75 had public advisories issued.[1] Picus Security summarizes Anthropic’s disclosures more broadly as “fewer than 1% of vulnerabilities found by Mythos were patched.”[5]

This gap between discovery and remediation is central to the tone of the update and of third‑party commentary: AI is massively increasing defensive visibility, but the ecosystem cannot yet absorb and fix this volume of high‑impact issues.[1][5][8]

Patching latency and security debt

Anthropic’s initial update gives a more granular view of how long it takes for Mythos‑found issues to be fixed:

Median time to patch a high/critical bug found by Mythos in open‑source projects is reported at around two weeks.[1]
However, at the time of reporting, the absolute number of patched issues remains small relative to the backlog: 75 patched vs. 530 reported high/critical, and >10,000 total discovered.[1][5]

This creates what security commentators are calling a newly exposed patching paradox:

On one hand, Project Glasswing demonstrates that AI can push defenders into a position where they see far more of their vulnerability surface, including many zero‑days, than before.[3][8][16 in your collection context]
On the other han

Key Insights

Claude Mythos Preview (Anthropic’s unreleased frontier model) now has a detailed public write‑up via the Project Glasswing “initial update”, showing it can autonomously find and exploit vulnerabilities at a level comparable to highly skilled professional security researchers.
In early Glasswing runs, Mythos Preview identified over 10,000 vulnerabilities, including 6,202 that it internally rates as high or critical severity across both open‑source and proprietary “globally important” software.
Despite this, patching is badly lagging: fewer than ~1% of vulnerabilities found by Mythos have been patched so far; on open source, only 75 of 530 high/critical bugs reported were fixed at the time of the report, with a typical patch latency of around two weeks.
The update confirms Mythos will not be generally released in its current preview form due to its offensive cyber capabilities; instead, access is being gated via Project Glasswing and $100m in free credits for vetted partners to use it defensively.
External analysts (e.g., Schneier, Cloudflare, independent security blogs) view Glasswing as both a real technical step change and a highly staged, capital‑intensive “demo environment” that exposes a structural problem: AI dramatically accelerates bug-finding, but the remediation pipeline and incentive structure have not caught up.

Key Entities & Research Links

COMPANYS

Anthropic
Context: Central AI company behind Claude Mythos Preview and Project Glasswing, making the decision to restrict release because of offensive cyber capability.
Cloudflare
Context: Security and infrastructure company mentioned as both a partner and an external commentator on Mythos and Glasswing.
Amazon Web Services
Context: Major cloud provider listed among the partners involved in Project Glasswing.
Microsoft
Context: Major technology company named as a Project Glasswing partner.
Google
Context: Major technology company listed among the consortium partners in Project Glasswing.
Apple
Context: Major technology company included among the partners supporting Project Glasswing.
JPMorganChase
Context: Financial institution named as a participant in the Glasswing partner ecosystem.
Broadcom
Context: Technology company listed among the security and infrastructure partners in Glasswing.
Cisco
Context: Networking and security company mentioned as part of the Glasswing partner group.
CrowdStrike
Context: Cybersecurity company cited among the Project Glasswing participants.
NVIDIA
Context: Technology company named as part of the Glasswing consortium.
Palo Alto Networks
Context: Cybersecurity company listed among the partners involved in Project Glasswing.
Picus Security
Context: Security analysis company cited for summarizing Anthropic's system card and reporting detailed findings.
OpenAI
Context: AI company mentioned in comparison as another lab withholding a similarly capable model from general release.

PRODUCTS

Claude Mythos Preview
Context: Anthropic's unreleased frontier model that is described as highly capable at autonomous vulnerability discovery and exploitation.

EVENTS

Project Glasswing
Context: Anthropic's controlled initiative for running Mythos against critical software and coordinating disclosure and patching with vetted partners.

PERSONS

Bruce Schneier
Context: Security commentator cited as framing Glasswing as a reactive step in an ongoing race against attackers.

ORGANIZATIONS

Linux Foundation
Context: Open-source steward referenced as a partner for securing critical software infrastructure.

Related Images

Project Glasswing: An initial update — Anthropic
Claude Mythos Preview \ red.anthropic.com
Securityweek
Anthropic Unveils 'Claude Mythos' - A Cybersecurity Breakthrough That Could Also Supercharge Attacks - SecurityWeek
Ethz
With Claude Mythos, a single hacker suddenly has a lot more ways to attack.” | ETH Zurich
Wandb
Claude Mythos Benchmark Scores | ml-news – Weights & Biases
Sanity
Everything You Need to Know About Claude Mythos

Sources & Citations

[2] FR/02/2026 Supervisory Toolkit for AI Use in Capital Markets
iosco.org
[1] What is Anthropic's Mythos AI model and why does it have the financial world in a panic?
financialpost.com
[3] Anthropic's Project Glasswing flags 10,000 critical software flaws - TechBriefly
techbriefly.com
[4] Anthropic Finds 10,000 Software Flaws In First Month Of Project Glasswing - Dataconomy
dataconomy.com
[5] Claude Mythos Preview has discovered 6,202 potential vulnerabilities estimated to be of 'high' severity or higher; an initial report on the high-performance AI, which is seeing moves toward use by Japanese banks, has been released. - GIGAZINE
gigazine.net
[6] The Real Context News AI Reading List: Anthropic Claude Mythos Glasswing Edition (#7) – Real Context News (RCN)
realcontextnews.com
[7] Anthropic Project Glasswing Targets Safer AI Agents Before They Go Rogue | HotHardware
hothardware.com
[8] Anthropic's Glasswing: 10,000+ Vulnerabilities Found in One Month, and the Patching Problem Has Never Been More Obvious
securityaffairs.com
[9] Anthropic says Mythos has already found more than 10,000 vulnerabilities - Engadget
engadget.com
[10] Anthropic's Claude Mythos Preview Uncovers 10,000+ 0-Days in Project Glasswing
cybersecuritynews.com
[11] Inside the White House’s AI reversal
politico.com
[12] PANW, NET: Anthropic’s Project Glasswing Update Casts Spotlight on Cybersecurity Stocks - TipRanks.com
tipranks.com
[13] Project Glasswing: Anthropic says Claude found 10,000 critical software flaws in a month
interestingengineering.com
[14] Project Glasswing: An initial update
anthropic.com
[15] Anthropic's Project Glasswing Finds 'More Than 10,000' Critical Bugs, Expands To Additional Partners - Benzinga
benzinga.com
[16] Project Glasswing: AI Finds 10,000+ Software Vulnerabilities | AIToolly
aitoolly.com
[17] Apple’s boldest iPhone yet: Quad-curved ‘Liquid Glass’ display set to eliminate bezel era
macdailynews.com
[18] Project Glasswing: Securing critical software for the AI era
anthropic.com
[19] Verizon Joins Anthropic’s Project Glasswing for AI Defence | AI Magazine
aimagazine.com
[20] IBM Brings Its Most Advanced AI-Powered Security Portfolio to Clients, and is Strengthened by Ongoing Project Glasswing Work
mcpressonline.com
[21] IBM Is The Newest Member Of Project Glasswing – How Does This Benefit The Software Industry?
stocktwits.com
[22] Cisco: Latest news and insights
networkworld.com
[23] The times they are a-changin’: Washington suddenly warms to regulating AI
fortune.com
[24] ⚡ Weekly Recap: Exchange 0-Day, npm Worm, Fake AI Repo, Cisco Exploit and More
thehackernews.com
[25] The iPhone and Mac security Apple spent 5 years building? AI broke it in 5 days.
cultofmac.com
[26] On Anthropics Mythos Preview And Project Glasswing
www.schneier.com
[27] Watch
www.youtube.com
[28] Item
news.ycombinator.com
[29] Anthropics Project Glasswing Paradox
www.picussecurity.com
[30] Project Glasswing The 10 Consequences Nobodys Writing Abo...
www.forrester.com
[31] Watch
www.youtube.com
[32] Cyber Frontier Models
blog.cloudflare.com
[33] Claude Mythos 2 Cybersecurity And
thezvi.substack.com

Continue Research: Access PowerIntell.AI Platform

Entity links are valid for 7 days and will automatically execute searches when clicked.

Shared via PowerIntell.AI